Robust monocular visual odometry for road vehicles using uncertain perspective projection
© Van Hamme et al.; licensee Springer. 2015
Received: 28 January 2015
Accepted: 1 April 2015
Published: 26 April 2015
Many emerging applications in the field of assisted and autonomous driving rely on accurate position information. Satellite-based positioning is not always sufficiently reliable and accurate for these tasks. Visual odometry can provide a solution to some of these shortcomings. Current systems mainly focus on the use of stereo cameras, which are impractical for large-scale application in consumer vehicles due to their reliance on accurate calibration. Existing monocular solutions on the other hand have significantly lower accuracy. In this paper, we present a novel monocular visual odometry method based on the robust tracking of features in the ground plane. The key concepts behind the method are the modeling of the uncertainty associated with the inverse perspective projection of image features and a parameter space voting scheme to find a consensus on the vehicle state among tracked features. Our approach differs from traditional visual odometry methods by applying 2D scene and motion constraints at the lowest level instead of solving for the 3D pose change. Evaluation both on the public KITTI benchmark and our own dataset show that this is a viable approach for visual odometry which outperforms basic 3D pose estimation due to the exploitation of the largely planar structure of road environments.
KeywordsVisual odometry Localization Computer vision SLAM
Visual odometry is an increasingly important research domain in the field of intelligent transportation systems. Many emerging and existing applications related to consumer vehicles rely on accurate position estimation. Some examples of these applications are navigation, lane assistance, collision warning, and avoidance. Traditionally, the positioning data for these applications is provided by satellite-based systems such as GPS, GLONASS, or GALILEO, sometimes augmented by closer-range communication (as in DGPS) or additional sensors scanning the local environment (e.g., a lane assist camera). However, the reliance of these applications on satellite navigation is a threat to their full-time availability. Due to the four-dimensional nature of the problem (3D positioning and time synchronization), signals from at least four satellites must be received in order to obtain a positional fix. It is well documented that in certain urban scenarios, large parts of the sky can be obscured by buildings or road infrastructure, making the reception of four satellites unlikely or even impossible for many seconds or even minutes . In these cases, satellite navigation systems cannot provide reliable position estimates.
In this context of assisted or autonomous driving, positioning solutions complementary to the satellite-based systems are needed. One such solution is visual odometry: the measurement of a vehicle’s trajectory using vehicle-mounted cameras. Visual odometry is closely related to simultaneous localisation and mapping (SLAM) in the field of robotics, but there are clear distinctions between the two. Whereas SLAM places equal emphasis on constructing a virtual map of the unknown environment as on positioning relative to that environment, visual odometry methods do not need to explicitly map the environment. The two problems remain strongly intertwined, as positioning relies on finding fixed points in the environment of the vehicle, but for visual odometry, the mapping itself is of little interest. In fact, in some cases (especially consumer automotive applications), some information about the environment may already be known (e.g., the local road network layout).
Visual odometry only provides relative positioning, i.e., positioning relative to an earlier visited reference point. As a consequence, estimation errors are cumulative, and visual odometry methods are therefore susceptible to drift. The greater the distance traveled from the last absolute reference point (e.g., the known GPS coordinates of the starting address), the greater the positional error can become. While this may appear to limit the use of visual odometry to a short distances, the drift error can be bounded by combination with additional passive sensors (e.g., a magnetic compass for dead reckoning) or a priori known information (e.g., the local road map) [2,3]. As such, visual odometry is still a prime candidate to supplement satellite navigation even for urban scenarios where signal reception may be unreliable for prolonged distances.
In the classical approach, visual odometry is a pose estimation problem in a calibrated setting. Given a camera with known intrinsic calibration parameters and the images of a scene captured from two unknown viewpoints, what is the relative camera pose between the two viewpoints? In this calibrated setting, visual odometry is achieved by estimating the essential matrix that relates the homogeneous image coordinates of the same world point in the two viewpoints up to a scale factor . A computationally efficient solution for this was published in the 1990s by Philip  and improved upon by Nistér . In Nistérs solution, a RANSAC algorithm evaluates sets of five correspondence points to find the best estimate for the essential matrix. This type of method is therefore called a five-point solver. The RANSAC algorithm is necessary to cope with outliers that will arise from erroneous feature matching and external circumstances (e.g., other traffic). The essential matrix can be decomposed into its rotation and translation components if necessary.
As a generalization of the classical setting, pose estimation can also be performed for uncalibrated cameras. In this case, the matrix that relates the image coordinates of the two viewpoints is called the fundamental matrix . It is estimated in a similar way to the essential matrix; however, more point correspondences are necessary. Methods of this type are called eight-point solvers. Even in the calibrated setting, there is merit in using an eight-point solver as it yields only one solution, while the five-point methods can produce up to ten valid solutions, requiring additional constraints to be evaluated.
The aforementioned methods for estimating the essential or fundamental matrix are affected by the problem of degenerate configurations. Two distinct cases of degeneracy arise: degeneracy in the motion, where the camera undergoes only rotation and little or no translation, and degeneracy in scene structure, where all or many of the points are coplanar. In both cases, the accuracy of the pose estimation will be severely degraded . This is an important drawback in real-world applications, where vehicles will often make small incremental motions and where the majority of the scene can consist of objects in or close to the ground plane. To remedy the problems of degeneracy, a stereo camera configuration is typically used, which allows for much better triangulation of the feature points even in the cases of motion or scene degeneracy.
Alternatives to fundamental matrix estimation for stereo systems have also been proposed based on triangulation through stereo disparity [9,10]. Typically, this class of algorithms first estimates approximate 3D coordinates from a stereo image pair and then links up feature tracks over multiple pairs to estimate camera motion.
Stereo camera setups however have significant downsides for consumer automotive applications. They are more expensive than a single-camera system and are more difficult to integrate into the car’s design. Additionally, they rely on very accurate calibration on account of the long observation distance to baseline width ratio . In the vibration and shock-prone environment of a car, it is generally accepted that long-term calibration stability cannot be guaranteed, and online recalibration methods have been proposed [12,13] in an effort to improve the applicability. Monocular solutions are inherently less susceptible to calibration drift, as fewer assumptions about the capture system’s geometry are made.
Monocular visual odometry algorithms that do not employ fundamental matrix estimation and are therefore not impacted by the aforementioned degeneracies have been proposed by Tardiff et al.  and Scaramuzza [15,16]. However, these methods are only demonstrated using an omnidirectional camera mounted atop the vehicle, which is not practical for application on consumer vehicles. More relevant is the work of Chandraker and Song . In this work, a five-point solver provides an initial triangulation of image points captured over five frames, after which new points are mapped to the known 3D structure and allow for four-point pose estimation. The output of the pose estimation is combined with continuous ground plane estimation in a data fusion framework, providing high accuracy as well as being unaffected by planar scene degeneracy. This proves the merit of combining different visual cues to improve the overall odometry accuracy. We expect this data fusion approach to be applied on other base odometry algorithms as well in the future.
Recently, a different approach to monocular visual odometry has emerged in literature, called direct or sometimes dense visual odometry. Instead of determining feature correspondences, these methods aim to recover the camera pose directly from the image data, by reconstructing a surface-based depth map for the image. While this approach is not new, only recently has it become tractable for real-time applications [18-21]. These methods perform very well for structure-rich indoor and outdoor environments, but to the best of our knowledge, their accuracy in sparsely structured open road scenes is yet to be examined.
In this work, we will present and evaluate a monocular visual odometry method that does not depend strongly on accurate camera calibration and does not suffer from degeneracy in case of small incremental motion or planar scene geometry. Furthermore, the method is suitable for any standard camera that views part of the road surface in front of or behind the vehicle. This is compatible with normal camera placement for other currently emerging automotive vision applications such as traffic sign recognition and obstacle detection. The method tracks ground plane features, taking into account the uncertainty of the camera viewing angle with relation to the ground plane. This allows us to exploit the inherently two-dimensional character of vehicle motion while still retaining some of the accuracy benefits of a fully three-dimensional approach. Additionally, the use of uncertainty margins relaxes the requirement of accurate camera calibration.
Two key components of the method provide robustness against the common problem of outliers: a feature matching method constrained by uncertainty zones and a Hough-like parameter space vote. The combination of these two mechanisms eliminates the need for a RANSAC scheme and speeds up computation, while still producing useful odometry for inlier ratios as low as 1:8 in real-world experiments.
This work is a continuation of the concept first introduced in our publication at IV2011  and tested in a real-world scenario at ITSC2012 . The contributions of this paper in addition to the prior work are an extended literature review comparing the different approaches to visual odometry and their relative merits, a proper analysis and justification of the proposed method’s underlying assumptions about vehicle dynamics, evaluation on two extended datasets, comparison against a reference method, qualitative assessment of the method’s main benefits, calibration sensitivity analysis, and quantification of the effect of non-planarity of the road surface.
The proposed method is shown to produce reliable visual odometry even for longer trajectories of several kilometers, and its accuracy compares favorably to the monocular instance of the eight-point solver of Geiger et al. , both on the public KITTI dataset  and on a 15-km dataset captured locally with the GrontMij mobile mapping vehicle. This proves that approaching visual odometry as a two-dimensional problem from the bottom up not only offers practical benefits with relation to robustness, execution speed, and calibration but also provides accuracy competitive with the traditional 3D pose estimation approach.
A detailed description of the method is given in Section 2. Details about the calibration procedure and sensitivity simulations are in Section 3. Experimental validation is provided in Section 4, with a discussion of the results in Section 5. Finally, conclusions about the viability of this type of monocular visual odometry are drawn in Section 6.
2 Algorithm description
The general structure of the method bears some resemblance to a Kalman filter in the sense that it uses a prior vehicle state estimate to predict current feature locations and then compares this prediction to current observations to calculate an updated vehicle state. However, because the accuracy of an observed feature depends strongly and nonlinearly on its position, novel strategies for prediction and update are implemented specific to this application.
In the prediction step, the previously estimated steering angle and velocity of the vehicle are used to define search regions in the ground plane where previously observed features may be found. The details of this prediction will be explained in Section 2.2.
The observation step is more complex, since we cannot measure ground plane coordinates directly using the perspective camera. In order to relate the image coordinates of features in the camera view to their ground plane positions, inverse perspective projection is performed. This inverse perspective projection is only determined for features originating from a plane with a known orientation relative to the camera. In other words, we can only determine the inverse perspective projection if we know the viewing angle of the camera to the ground plane. This angle, however, is not static. The suspension of the vehicle creates variability in the camera pose, and this translates to uncertainty on the inverse perspective transform. We will take into account this uncertainty and define plausible regions of ground plane coordinates for each feature point. This process will be explained in more detail in Section 2.1.
In the update step, the predictions of feature locations are compared to the observations. In our method, this corresponds to matching the predicted search regions for previously seen features with the uncertainty regions pertaining to the inverse perspective projection of currently seen features. From these matches, a consensus is drawn to update the vehicle state. In this match-and-update step, several mechanisms will ensure robustness to outliers in the input data. This is explained in Section 2.3.
Finally, the vehicle trajectory can be calculated from the consecutive vehicle states.
2.1 Inverse perspective projection
To describe the inverse perspective projection, we first need to define the forward perspective projection that describes the image capturing process. Much of this section follows the standard model for the projective camera as described by Hartley and Zissermann . This model describes a transformation from the 3D world axes to the 2D image axes of the captured video frame. This transformation consists of two steps.
in which [R | t] is the rotation matrix R that aligns the world axes with the camera axes, augmented by the 3D translation vector t between their origins.
By affixing the 3D world axes to the vehicle, the matrix [R | t] is made independent of vehicle position, and R and t can be determined by extrinsic calibration (e.g., using Zhang  or Miksch et al. ). Vehicle motion will now manifest itself as a change in coordinates of the feature points corresponding to static objects in the real world.
is the projection matrix consisting of the horizontal and vertical focal lengths α x and α y and the coordinates of the principal point (i.e., the projection of points on the Z axis) (x 0,y 0). These intrinsic parameters can easily be estimated using a standard camera calibration method (e.g., Bouguet ). The factor w serves to compress the 3D space onto the sensor plane by scaling the X and Y coordinates by the inverse of the Z coordinate.
in which R XY is the submatrix of R obtained by omitting the third column. The inverse perspective projection that maps homogeneous image coordinates back onto homogeneous 2D world ground plane coordinates is then the inverse of C[R XY |t]. Inverse perspective projection is sometimes also referred to as backprojection.
An important remark with respect to backprojection is that the calculated transform is only valid for features corresponding to objects in the ground plane. However, there is no easy way to discern the Z coordinate of a feature in the camera view. We therefore have no choice but to apply the backprojection to any features we detect in the camera image and sort out the above-ground features in a higher level reasoning step.
Throughout this discourse, we assumed the matrix [R|t] to be known from extrinsic calibration. In practice, however, the camera coordinate system is not rigidly affixed to the axles of the vehicle. Instead, the camera is attached to the body of the vehicle, which has a variable pose with relation to the axles on account of the suspension travel. The matrix [R|t] is therefore no longer static but changes somewhat as the vehicle moves along. It is important to take this pose variation into account as it directly affects the estimated ground plane coordinates of all features.
Since we have no reliable way of measuring the attitude of the vehicle, we model the uncertainty on the inverse perspective projection arising from this attitude. To this end, we determine realistic limits on the suspension travel during normal driving and calculate the inverse perspective transforms corresponding to these limits. This yields a region of possible ground plane coordinates for each feature detected in the perspective view.
A typical road vehicle experiences in the range of 100 to 150 mm total suspension travel measured at each wheel. With a track width of 1.4 to 1.5 m, this could theoretically give rise to approximately 10° of lateral roll. Considering a wheelbase of 2.7 m on average, the maximum pitch is approximately 5°. However, these limits would be very hard to achieve in practice even with extremely aggressive driving, as the vehicle will tend to break traction first. In typical town driving, more representative values for maximum roll and pitch are respectively 2° and 1° either side of the level position. For highway driving, the expected angles are even smaller.
A final remark concerns lens distortion. The above pinhole camera model does not take any distortion into account. In order for this model to be a good approximation, the distortion must either be small or corrected in pre-processing. Especially when using wider angle lenses, it is recommended to estimate radial distortion parameters (e.g., using Bouguet ). As these parameters are largely stable for a lens with fixed focal length, this kind of calibration does not need to be recurrent. For consumer vehicles, this means the distortion parameters can be determined at the factory or even specified by the supplier of the optics.
2.2 Feature matching
Our method will use the rotation and velocity estimated in the previous timestep to predict such a search region for all currently tracked features. The prediction uncertainty region for each feature is closely approximated by the quadrilateral defined by the predictions corresponding to extremal combinations of steering angle and velocity.
The theoretical maximum change in vehicle speed corresponds to an emergency stop and is around 10 m/s2. Again though, typical values during normal driving are much less extreme. Maurya and Bokare  measured maximum deceleration for cars in hard braking from motorway speeds to be 1.71 m/s 2. For trucks, this value is reduced to 0.88 m/s 2. On our own data, obtained using a family sedan and a van, we observed maximum deceleration to be under 1.5m/s 2. The maximum rate of acceleration of a normal road vehicle is significantly lower than the maximum rate of deceleration ; therefore, we will assume acceleration in normal circumstances to be under 1.5 m/s 2 as well.
In a road driving context, location-based matching is generally preferable to appearance-based matching, as it is to be expected that many features on the road surface will have the same general appearance and the number of possible matches to be evaluated will therefore be much higher than when using a location-based approach.
Another benefit of location-based matching is that it will produce fewer spurious matches in the event of other moving objects being present in the camera view. Features detected on this moving object will, in general, not have observation uncertainty regions that consistently overlap with prediction uncertainty regions because the relative motion of the object does not comply to the constraints of the Ackermann model. This is a significant advantage compared to an appearance-based matcher, which will tend to match a large amount of features exhibiting a consistent motion pattern which may be hard to discern from the motion pattern of road surface features. Similarly, our matching principle will not generally produce matches for features that originate from a point significantly above the ground plane, as these features will exhibit exaggerated motion compared to actual ground plane features and therefore fall outside of the prediction uncertainty regions.
2.3 Odometry estimation
Outliers may be caused by accidental matches of moving objects in the scene, by overlapping of uncertainty regions of multiple features with the same search region or vice versa. Also, some of the matches may be inliers but still unreliable for calculating odometry, on account of them not originating from the ground plane. Features on slightly elevated curbs, for example, will generally match, though their uncertainty regions are inaccurate. A more in-depth analysis of the degeneracy that occurs when many features are in a slightly elevated plane is presented in the Appendix.
In a traditional visual odometry framework, the calculation of relative pose change from unreliable feature matches consists of a RANSAC scheme to sort inliers from outliers and find the best supported motion hypothesis. However, RANSAC offers few advantages in our case, as the majority of the outliers have already been eliminated by the location-based matching, and any remaining outliers are difficult to identify due to the uncertainty of the observed feature coordinates associated with both inliers and outliers.
Instead of relying on RANSAC, our method employs a parameter space voting approach. This integrates well with our uncertainty regions and will allow us to easily find a consensus among the matches. Let us revisit the prediction uncertainty regions for each feature (as seen in Figure 7). The edges of these predicted regions correspond to the limits of change the driver can affect on the vehicle state, while the center of the regions corresponds to an unchanged vehicle state. As such, each prediction uncertainty region represents the same patch in rotation-velocity parameter space, centered around the last estimates for rotation and velocity. When an observation uncertainty region of one of the current features overlaps with part of one of the predicted regions, the overlap expresses a vote of this feature on a part of the rotation-velocity parameter space patch. For example, if the observation uncertainty region overlaps with the left side of the prediction uncertainty region, this corresponds to an increased likeliness that the vehicle has turned further to the right or less to the left than in the previous inter-frame interval.
In practice, the discretized nature of the sum image and the limited number of features means that the location of the peak intensity is quite sensitive to noise (e.g., a single feature that has shifted by one pixel in the camera image between frames could have a significant impact on the location of the maximum). In order to reduce this noise sensitivity, we will not locate the absolute maximum but the center of gravity of the area of highest intensity. We define this area as the region in which the values exceed a fraction (typically 70%) of the absolute maximum. The center of gravity calculation is essentially an averaging mechanism and therefore reduces noise sensitivity.
The location of the center of gravity can be easily related to its corresponding values in rotation-velocity parameter space, which define the current vehicle state. When the vehicle state is known in every inter-frame interval, the complete estimated trajectory of the vehicle can be reconstructed using the circular motion model described in Section 2.2.
An important remark should be made about the accuracy of this estimation. Due to the uncertain nature of the observations (i.e., the significant size of the backprojected regions) and the limited sampling density in the parameter space, the immediate frame-to-frame estimate is of relatively low accuracy. The uncertainty on the pitch and roll angles prevent us from refining this estimate further through a closed-form calculation (e.g., a least squares solution). However, our method is self-correcting in the sense that an estimation error will result in a prediction for the next frame that is biased in the direction of the error. The observations will then accumulate votes in an area offset in the opposite direction of the prediction bias. As a consequence, the consecutive estimation errors will not accumulate but compensate each other instead. Therefore, the cumulative vehicle state over several frames will prove more accurate than the fuzzy nature of the data suggests. To illustrate this point, consider the simplified example of overestimating the velocity at time t as 0.45 m/frame while the real velocity is just 0.4 m/frame. This overestimation of the velocity is equivalent to a misestimation of the actual feature positions from the fuzzy data by 0.5 m. The prediction for time t+1 will assume a constant velocity of 0.45 m/frame and use the misestimated actual feature coordinates as a starting point. The centers of the prediction uncertainty regions for time t+1 will therefore end up at a distance of 0.10 m to the actual feature coordinates. When the actual velocity of the vehicle at time t+1 is again 0.4 m/frame, the observation uncertainty regions will then each be centered on a pixel corresponding to 0.10 m above the center of a prediction uncertainty region. The method will then correct the estimate for the second state to 0.35 m/frame, and the average estimated velocity over two states will be accurate. This safety mechanism will only mitigate single-frame estimation errors; in case of continuously poor feature matching, errors may still accumulate.
Another remark should be made about the area of the rotation-velocity parameter space that falls outside of the predicted boundaries. This area is not taken into account for the parameter space vote. By cutting off the parts of the observation uncertainty regions at these boundaries and not taking them into account for the center of gravity calculation, we introduce a slight bias towards the center of the rotation-velocity parameter space patch. This bias is not a problem; as explained above, it is automatically corrected for in the next estimation step as long as the parameter space boundaries are chosen sufficiently wide to accomodate this extra frame-to-frame variability. The estimation bias towards the unchanged vehicle state hypothesis also limits the error caused by low feature quality. In such cases of low feature quality (caused, for example, by excessive camera vibration), the sum image will degrade into noise, and it is beneficial to overall robustness to assume a stable vehicle state in this case that can be corrected when feature quality improves.
As a final step in the odometry method, the feature tracks need to be updated. In the discussion so far, we have assumed the ground plane coordinates of each tracked feature at the previous timestep to be known. Due to the uncertain inverse perspective projection, however, these coordinates cannot be determined exactly. As a best estimate for the ground plane position corresponding to the a feature observation, we will use the centroid of its observation uncertainty region. The estimated vehicle state (rotation and velocity) is used to update this centroid at each timestep. Additionally, for any feature detected in the camera image that did not match any prediction uncertainty regions, a new track is initiated with the centroid of its observation uncertainty region as starting coordinates. Finally, tracked features which have not matched with any observations for a number of consecutive frames (typically chosen between three and five) are discarded.
In this section, we will describe how the extrinsic calibration can be determined. The extrinsic calibration is contained in the matrix [R XY | t] in Equation 5. R XY is a submatrix of the 3D rotation matrix R that describes the rotation between the world axes and the camera axes. Homography-based methods can be found in literature to estimate this rotation matrix, notably the work of Miksch et al. , who determine the rotation matrix online without using odometry data or known scene geometry, but with known camera height, using inter-frame feature correspondences on the ground plane.
The extrinsic rotation matrix can also be estimated iteratively from a single image using known scene geometry, e.g., the known dimensions of a rectangular parking space. When the rotation matrix R is known, the translation vector t can be easily measured in vehicle axes with a tape measure and plumb rule and then rotated into camera axes using R −1.
From the error graphs and reconstructed trajectories, we can see that three parameters are especially important for translation accuracy. The greatest translation error occurs in the case of misestimated pitch (Figure 13, third from top), with a 1° error resulting in an overestimation of travel distance by 13%. A 10-cm vertical or longitudinal offset both result in a translation error of around 6% (Figure 13, bottom two). Inaccuracies in the other parameters yield much smaller translation errors.
In terms of rotation error, the single most important parameter is the heading angle (Figure 13, fourth from top). An error of 1° in this parameter results in a rotation error of 0.11°/m. The roll angle is the second most important influence on rotation accuracy (Figure 13, second from top), with a 1° roll misestimation resulting in a 0.04°/m error. The other calibration parameters have smaller effects on rotation error.
We can conclude that the proposed method is most sensitive to pitch and heading angle, followed by roll angle, vertical offset, and longitudinal offset. Generally, the offsets are easy to measure in practice, and an error of 10 cm is not likely. Estimation of the extrinsic rotation angles is more prone to inaccuracies. However, as each of the three angles has a different effect on the evolution of the error on our simulated trajectory, it is easy to identify an error in one angle. In practice, this can be done by driving along a known section of road featuring at least one straight section and a bend in each direction and comparing the odometry result to the known ground truth. A roll error causes a significant rotation bias on the straight sections, but not in the bends. This property can be used to refine the roll estimate. A heading error causes a constant bias regardless of road curvature, making it easy to identify and correct as well. Finally, a pitch error results in over- or underestimation of rotation in bends only and a significant constant bias on translation. If the longitudinal offset (which has similar effects) is reliable, the translation error by itself can be used to correct the pitch angle. Although these principles have already been used to manually refine the calibration estimate in some of our experiments, the automation of the process for mixed calibration errors remains future work.
The proposed method was evaluated on two datasets and compared to the monocular eight-point solver by Geiger et al. . The implementation is provided online by the authors. In the KITTI odometry benchmark, three monocular methods currently outperform this standard eight-point solver. The highest ranked method, by Chandraker and Song , uses a standard five-point solver as one of the base components, combined with ground plane estimation and scene structure propagation. The second method, called windowed structure from motion (W-SFM) lists no publication but is described as using a five-point solver and bundle adjustment. The third method is the eight-point solver of Geiger et al. combined with the ground plane estimation proposed by Chandraker and Song. All three methods employ either a five-point or eight-point solver to perform initial 3D pose estimation. The aim of this research is to prove that our 2D approach is a viable alternative to traditional 3D pose estimation for visual odometry. We have therefore chosen the basic eight-point solver as reference method. The potential improvements afforded by bundle adjustment and more precise ground plane estimation for the proposed method are to be explored in future work.
It can be easily verified that Δ r corresponds to the heading angle difference between Q and Q ′ when the rotation is limited to the Z-axis.
The translation and rotation errors are calculated on all subsegments of the ground truth trajectory of length 100, 200 …800 m. The errors is averaged per segment length. Translation error is expressed as a percentage of segment length, while rotation error is expressed in °/m.
The proposed method only estimates rotations along one axis (the normal of the ground plane) and does not measure elevation change, while the evaluation considers full 3D poses and elevation change. The KITTI dataset contains several sequences captured on hilly roads, and we can expect the proposed method to be at a slight disadvantage in this benchmark as a result, while for real-world navigation-related applications, the elevation changes are largely irrelevant due to the planar nature of common map data.
Both methods were able to process the data faster than real-time on a desktop computer (Intel Core i5 3.40 GHz ×4), with the proposed method significantly outperforming the eight-point method (86.4 vs. 17.2 fps). The feature detection step in the proposed method is implemented to make use of multi-core systems (in this case running on four cores), while the remainder of the processing is single threaded. The method of Geiger et al. runs completely single threaded.
The slightly downward pitch of the camera in these video sequences is considered slightly better for the proposed method, as it offers a denser coverage of the nearby road plane. A second difference with the KITTI set is the reduced horizontal angle of view of the camera. In the KITTI set, the aspect ratio is about 3.3, significantly wider than the standard 1.78 widescreen ratio of the HD camera used for the Diepenbeek/Hasselt set. This narrower field of view means that average feature displacement for a given speed is reduced (since features off to the sides have the greatest displacements), and therefore, the triangulation accuracy is also expected to be slightly lower.
Overall, the proposed method is markedly better than the method of Geiger et al. in both metrics on both datasets.
Summary of mean errors of both methods on both datasets
Rot.err. (° /m)
Geiger et al.
Geiger et al.
The results obtained on both datasets clearly illustrate the main advantage of the proposed method over the eight-point solver, namely, better recovery of scale. In several of the sequences, the eight-point solver significantly misestimates the length of one or more straight segments (e.g., the final section in the left plot of Figure 16). This is due to an inherent weakness in the monocular pose estimation. Due to the projective nature of the camera, the translation can only be recovered from the fundamental matrix up to a scale factor. As was noted in Kitt et al. , this scale factor is susceptible to drift. Scale drift is remedied in Geiger’s method by relating the triangulation of points to a known length in the scene, specifically the height of the camera above the ground plane (which is assumed constant). The results both on the KITTI and the Diepenbeek/Hasselt datasets clearly show that this corrected scale is less accurate than the scale obtained by our robust tracking of ground plane features. The fact that Geiger et al. are better able to recover the scale on the Diepenbeek/Hasselt dataset than on the KITTI dataset further corroborates this explanation: in the Diepenbeek/Hasselt set, the camera is placed significantly higher above the ground plane, which means that similar absolute errors in the estimation of the ground plane have a smaller effect when divided by the longer fixed distance.
An important trend can be observed in the results of both methods. Rotational error decreases with increasing segment length. We may conclude from this that there is some noise present on the immediate poses estimated by both methods, which averages to zero over many estimations.
The effect of vehicle speed on the translation and rotation errors is less clear from the plots, as the two datasets show slightly different trends. The high errors of both methods for low speeds on the Diepenbeek/Hasselt dataset can be explained by the fact that the low speeds mostly prevail in the busy city center, where the presence of other traffic degrades the results somewhat. In the KITTI dataset, this correlation between speed and traffic density is not present, and as such, the proposed method does not exhibit significant sensitivity to vehicle speed.
For the proposed method, we observed that meaningful vehicle states were produced for inlier ratios as low as 1:8, counted as features generating uncertainty region overlap divided by total feature count. This proves the efficacy of the location-based matching and the parameter space voting to extract odometry from noisy and unstable features. Below inlier ratios of 1:8, assuming an unchanged vehicle state proved better than using the state estimation, this 1:8 threshold on inlier ratio was added to the method.
A remarkable difference between the results of the two datasets is that on the KITTI sequences, the translation error of both methods is decreasing for increasing segment length, while on the Diepenbeek/Hasselt data, the opposite is true. This can be explained by the fact that the vehicle’s trajectory in the KITTI set is in general more compact; many of the sequences contain multiple loops and the starting and ending position are often close to each other. The Diepenbeek trajectories are less circular in nature. It can easily be seen that having loops or u-turns in a segment will reduce the absolute error over this segment compared to a segment of the same length but with a larger offset between start and end position. We consider the Diepenbeek/Hasselt set to be more representative of a typical car journey as it is a 15-km two-way travel from Diepenbeek to Hasselt and back, rather than an artificial data acquisition trajectory with the aim of covering as many streets and turns as possible in a short time and small area.
Looking at the estimated trajectories in more detail, we see that the proposed method has a significant rotational bias on some segments. One examples can be seen in the bottom left plot of Figure 19. This is due to the non-planarity of the road environment in those segments. A more in-depth analysis of these situations is given in the Appendix. The method of Geiger et al. does not suffer from this flaw. It is therefore to be expected that on long, straight roads, the eight-point solver will provide more reliable heading estimation. As both of the evaluated datasets feature many turns in quick succession due to the suburban environment, this is not readily apparent from the performance numbers.
Overall, we can observe that the proposed method is often better than the eight-point method at recovering macro-maneuvres present in the trajectory: at intersections and roundabouts, the eight-point solver sometimes fails to accurately estimate the large changes in heading. An example can be seen in the top right image of Figure 19: the method of Geiger et al. misses most of the roundabout. The ability to correctly estimate big maneuvres is especially important for the integration with offline map data, as maneuvres are generally more reliable clues for map matching than gentle curves and straights. The concept of map matching as a mechanism to eliminate error accumulation has already been proven [2,23].
We have proposed a monocular visual odometry algorithm that uses planar tracking of features rather than traditional 3D pose estimation. It is demonstrated that in a typical monocular setting, the method has a significant performance advantage over traditional fundamental matrix estimation.
The proposed method is applicable to both forward- and rearward-facing cameras and is proven to work well on camera pitch angles ranging from horizontal (zero pitch) to 20° downwards. We may assume that similar or higher performance will be achieved as long as the camera view covers the ground plane up to a distance of approximately 12 m (this is the nearest cutoff point used for feature detection). The camera height over the two datasets also differs significantly (1.5 and 2.7 m) so we may conclude that our method is applicable to a wide range of vehicles and camera mount points.
The improvement over the eight-point method brings the translation error of the proposed method in a difficult but representative real-world scenario down to around 15 per 200 m traveled on average. Several techniques have already successfully been applied to improve the results of the standard eight-point solver, such as bundle adjustment and ground plane normal estimation . In future work, we expect these techniques to further improve the proposed method as well.
As visual odometry will typically be only one component in a mixed-data system (e.g., coupled with an offline map and magnetic compass), it is our opinion that the performance improvement of the proposed method over the standard eight-point solver is significant and can make a large contribution to a navigation system which does not depend on any outside communication.
In our work, we have made the assumption that the road surface is planar. However, in reality, there are two important scenarios in which this assumption is violated, but in such a way that the outlier removal mechanisms of the proposed method are ineffective. We therefore call these scenarios degenerate configurations for the proposed method.
These errors are significantly mitigated when feature points are present in the center section of the road as well. In this case, the consensus is still formed primarily by planar features, and the elevated features have a smaller influence. A simulated video frame for this situation is shown in Figure 22 (right) and the resulting errors in Figure 23 (bottom). The errors in this case are insignificant at only 0.004°/m for rotation and 0.1% for translation.
The second scenario is that of a road with a crown, i.e., a road with a crown in which the center line is higher than the edges to improve water drain properties. On bidirectional single-lane or two-lane roads, this is a common property. On highways or unidirectional roads, a crownless sloped design is the norm. In the latter case, the planarity assumption holds from the point of view of the vehicle, as the axles of the vehicle remain parallel to the entire span of road surface. In the case of a crowned road, however, the two sides of the road are in different planes and this will cause the inverse perspective transform to be inaccurate for part of the features when the vehicle is driving on one side of the center line or for all of the features when the vehicle is driving over the center line.
To quantify the deterioration of the odometry result in these two cases, two simulations were performed similar to those mentioned in Section 3. In the first simulation, points beyond the left side of the vehicle were sloped downwards with a 4% slope. This corresponds to what can be expected when a vehicle drives on the right lane of a two-lane road crowned at the typical recommended slope of 2% . The odometry errors were only evaluated on the straight sections, as superelevation (i.e., a single-slope, banked turn) is generally used in bends instead of a crowned design. In the worst-case scenario, with no feature points in the center section, the rotation error was 0.013°/m and the translation error −0.5%. These errors are an order of magnitude smaller than those caused by the kerb scenario or in the calibration experiments. We may conclude that for a typical two-lane road, the crown does not cause significant errors in the odometry estimation.
We may conclude that while the outlier removal mechanisms in the proposed method cannot completely avoid errors caused by non-planarity of the road, the impact of these errors in typically occurring road geometries is low. In the worst-case scenarios, performance is still acceptable, although the non-planarity may become the dominant error source.
This research was made possible through iMinds, an interdisciplinary research institute founded by the Flemish Government.
- P Huang, Y Pi, Urban environment solutions to GPS signal near-far effect. IEEE Aerosp. Electron. Syst. Mag. 26(5), 18–27 (2011). doi:10.1109/MAES.2011.5871387.View ArticleGoogle Scholar
- M Brubaker, A Geiger, R Urtasun, in Conf. Computer Vision and Pattern Recognition. Lost! Leveraging the crowd for probabilistic visual self-localization (IEEEPiscataway, NJ, USA, 2013).Google Scholar
- I Parra Alonso, D Fernandez Llorca, M Gavilan, S Alvarez Pardo, MA Garcia-Garrido, L Vlacic, MA Sotelo, Accurate global localization using visual odometry and digital maps on urban environments. IEEE Trans. Intell. Transp. Syst. 13(4), 1535–1545 (2012). doi:10.1109/TITS.2012.2193569.View ArticleGoogle Scholar
- HC Longuet-Higgins, A computer algorithm for reconstructing a scene from two projections. Nature. 293, 133–135 (1981). doi:10.1038/293133a0.View ArticleGoogle Scholar
- J Philip, A non-iterative algorithm for determining all essential matrices corresponding to five point pairs. Photogrammetric Record. 15(88), 589–599 (1996).View ArticleGoogle Scholar
- D Nister, An efficient solution to the five-point relative pose problem. IEEE Trans. Pattern Anal. Mach. Intell. 26(6), 756–770 (2004). doi:10.1109/TPAMI.2004.17.View ArticleGoogle Scholar
- R Hartley, A Zisserman, Computation of the fundamental matrix F, Mult. View Geometry Compu. Vis (Cambridge Univ. Press, Cambridge, UK, 2004).View ArticleGoogle Scholar
- PHS Torr, AW Fitzgibbon, A Zisserman, The problem of degeneracy in structure and motion recovery from uncalibrated image sequences. Int. J. Comput. Vis. 32(1), 27–44 (1999).View ArticleGoogle Scholar
- F Bellavia, M Fanfani, F Pazzaglia, C Colombo, in Image Analysis and Processing ICIAP 2013. Lecture Notes in Computer Science, 8156, ed. by A Petrosino. Robust selective stereo slam without loop closure and bundle adjustment (SpringerNew York, NY, USA, 2013), pp. 462–471.View ArticleGoogle Scholar
- H Badino, A Yamamoto, T Kanade, in International Workshop on Computer Vision for Autonomous Driving @ ICCV. Visual odometry by multi-frame feature integration (IEEEPiscataway, NJ, USA, 2013), pp. 222–229.Google Scholar
- TD A. Bodis-Szomoru, Z Fazekas, in Stereo Vision. Calibration and sensitivity analysis of a stereo vision-based driver assistance system (IntechRijeka, Croatia, 2008), pp. 1–27.Google Scholar
- T Dang, C Hoffmann, C Stiller, Continuous stereo self-calibration by camera parameter tracking. IEEE Trans. Image Process. 18(7), 1536–1550 (2009). doi:10.1109/TIP.2009.2017824.View ArticleMathSciNetGoogle Scholar
- M Miksch, B Yang, K Zimmermann, in Intelligent Vehicles Symposium (IV), 2010 IEEE. Automatic extrinsic camera self-calibration based on homography and epipolar geometry (IEEEPiscataway, NJ, USA, 2010), pp. 832–839. doi:10.1109/IVS.2010.5548048.View ArticleGoogle Scholar
- J-P Tardif, Y Pavlidis, K Daniilidis, in Intelligent Robots and Systems, 2008. IROS 2008. IEEE/RSJ International Conference On. Monocular visual odometry in urban environments using an omnidirectional camera (IEEEPiscataway, NJ, USA, 2008), pp. 2531–2538. doi:10.1109/IROS.2008.4651205.View ArticleGoogle Scholar
- D Scaramuzza, Performance evaluation of 1-point-ransac visual odometry. J. Field Robot. 28(5), 792–811 (2011). doi:10.1002/rob.20411.View ArticleGoogle Scholar
- D Scaramuzza, R Siegwart, Appearance-guided monocular omnidirectional visual odometry for outdoor ground vehicles. IEEE Trans. Robot. 24(5), 1015–1026 (2008). doi:10.1109/TRO.2008.2004490.View ArticleGoogle Scholar
- S Song, M Chandraker, in CVPR, Columbus, Ohio, USA. Robust scale estimation in real-time monocular SFM for autonomous driving (IEEEPiscataway, NJ, USA, 2014).Google Scholar
- J Stühmer, S Gumhold, D Cremers. Real-time dense geometry from a handheld camera, Darmstadt, Germany (SpringerNew York, NY, USA, 2010), pp. 11–20.Google Scholar
- A Wendel, M Maurer, G Graber, T Pock, H Bischof, in Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference On. Dense reconstruction on-the-fly (IEEEPiscataway, NJ, USA, 2012), pp. 1450–1457. doi:10.1109/CVPR.2012.6247833.View ArticleGoogle Scholar
- J Engel, J Sturm, D Cremers, in Computer Vision (ICCV), 2013 IEEE International Conference On. Semi-dense visual odometry for a monocular camera (IEEEPiscataway, NJ, USA, 2013), pp. 1449–1456. doi:10.1109/ICCV.2013.183.View ArticleGoogle Scholar
- J Engel, T Schops, D Cremers, in Computer Vision ECCV 2014, Lecture Notes in Computer Science, 8690, ed. by D Fleet, T Pajdla, B Schiele, and T Tuytelaars. LSD-SLAM: large-scale direct monocular slam (Springer, 2014), pp. 834–849.Google Scholar
- D Van Hamme, P Veelaert, W Philips, in Intelligent Vehicles Symposium (IV), 2011. Robust monocular visual odometry by uncertainty voting (IEEEPiscataway, NJ, USA, 2011), pp. 643–647. doi:10.1109/IVS.2011.5940453.View ArticleGoogle Scholar
- D Van Hamme, P Veelaert, W Philips, in Intelligent Transportation Systems (ITSC), 2012 15th International IEEE Conference On. Communicationless navigation through robust visual odometry (IEEEPiscataway, NJ, USA, 2012), pp. 1555–1560. doi:10.1109/ITSC.2012.6338668.View ArticleGoogle Scholar
- A Geiger, LIBVISO2: C++ Library for Visual Odometry 2. http://www.cvlibs.net/software/libviso. Accessed 21 May 2014.
- A Geiger, CSP Lenz, R Urtasun, The KITTI Vision Benchmark Suite. http://www.cvlibs.net/datasets/kitti/eval_odometry.php. Accessed 10 Jan 2015.
- Z Zhang, A flexible new technique for camera calibration. IEEE Trans. Pattern Anal. Mach. Intell. 22(11), 1330–1334 (2000). doi:10.1109/34.888718.View ArticleGoogle Scholar
- J Bouguet, Visual methods for three-dimensional modeling (PhD thesis, California Institute of Technology, 1999).Google Scholar
- E Davin, Parameter identification of a linear single track vehicle model (Technical report, Technical University of Eindhoven, 2011).Google Scholar
- AK Maurya, PS Bokare, Study of deceleration behaviour of different vehicle types. Int. J. Traffic Transp. Eng. 2(3), 253–270 (2012).View ArticleGoogle Scholar
- G Long, Acceleration characteristics of starting vehicles (Technical report, Transportation Research Board, 2000).Google Scholar
- BM Kitt, J Rehder, AD Chambers, M Schonbein, H Lategahn, S Singh, in Proc. European Conference on Mobile Robots. Monocular visual odometry using a planar road model to solve scale ambiguity (IEEEPiscataway, NJ, USA, 2011).Google Scholar
- JA Rosenow, in MnDOT Road Design Manual. Cross sections, (2012).Google Scholar
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.