I Introduction
Robust and accurate perception of the surrounding environment is critical for downstream modules of an autonomous driving system, such as prediction and planning. In practice, perceptual errors result in braking and swerving maneuvers that are unsafe and uncomfortable. Most autonomous driving systems utilize a detectthentrack approach to perceive the state of objects in the environment [1]. This approach has strongly benefited from recent advancements in 3D object detection [2, 3, 4, 5, 6] and state estimation [7, 8, 9]. However, making this architecture robust is an open challenge, as it relies on the geometric consistency of object detections over time. In particular, detectthentrack must handle (at least) the following errors:

false negatives, i.e. missed detections;

false positives, i.e. hallucinated objects [10];

out of ontology objects, i.e. object categories not labeled at training time and hence not detected or recognized by the detector, e.g., road debris, wild animals;

incorrect motion estimation due to large viewpoint changes, occlusions, and temporally inconsistent detections in dynamic scenes;

incorrect associations due to compounding errors in tracking or poor state initialization.
To tackle these challenges, we present a LIDARbased scene motion estimator that is decoupled
from object detection and thus complementary. Our method takes two consecutive full LIDAR point cloud sweeps as input. Each LIDAR sweep is encoded into a discretized 2D BeV representation of learned feature vectors, a.k.a. “pillars”
[6]. Then, we learn an optical flow network, adapted from Sun et al. [11], to locally match pillars between the two consecutive BeV feature grids. The whole architecture is learned endtoend and the final output is a 2D flow vector for each grid cell.Our approach relies on a 2D BeV representation over a 3D or projective representation (depth image) for multiple reasons. First, for autonomous driving, we primarily care about motion occurring on the road and adjacent surfaces, especially for motion planning. Second, this Euclidean representation allows us to design the network architecture to leverage spatial priors on relative scene motion. Finally, a 2D representation is more computational efficient compared to volumetric approaches and facilitates the sharing of representations with an object detector running in parallel to our objectagnostic flow network.
Figure 1 shows an example of our flow estimation results in bird’s eye view, where the moving vehicles are accurately inferred by the estimated flow (colorful regions). Our main contributions can be summarized as follows.

We propose an endtoend method, PillarFlow, to effectively estimate dense local 2D motion in a LIDAR BeV representation. Our deep net can leverage contextual knowledge of the scene and generalize to properly estimate the motion of unseen object types.

We integrate and leverage our proposed 2D BeV flow estimation to improve object tracking performance on both a public and an internal dataset.

We demonstrate the computational efficiency, robustness, and practical use of our approach by integrating it in a realworld autonomous driving platform operating in challenging urban conditions.
The rest of the paper is organized as follows. Section II gives a brief overview of related work. Section III depicts the proposed system and network architectures. We present our experimental results on public and inhouse datasets in Section IV, followed by a realworld integration and indepth analysis in an autonomous car in Section V. We report our conclusions in Section VI.
Ii Related Work
Iia Scene Flow Estimation
To estimate motion in the surrounding world, many approaches have been developed to estimate scene flow directly from LIDAR sweeps. For instance, Dewan et al. [13] formulate rigid scene flow estimation as an energy minimization problem using SHOT feature descriptors. Ushani et al. [14] propose a learningbased flow estimation that trains an encoding network to extract binary vector features from 3D points. Instead, we use pillar feature for better representation. Other works [15, 16] rely on range images projected from LIDARs for flow estimation in projective view.
A common alternative for scene flow estimation is to use unstructured pointbased representations. Gu et al. [17] propose an endtoend deep network to fuse features from unstructured point clouds from two consecutive LIDAR sweeps. Wang et al. [18] propose a parametric continuous convolution layer for nongridstructured data, and demonstrate the application in point cloud segmentation and LiDAR motion estimation. Liu et al. [19] introduce FlowNet3D, which builds on PointNet++ [20], leveraging a flow embedding layer to fuse two consecutive LIDAR sweeps. Wang et al. [21]
improve on FlowNet3D by using additional geometric loss functions beyond the L2 distance (Point to Plane and Cosine Distance). Behl
et al. [22]propose PointFlowNet to jointly train the tasks of 3D scene flow, rigid motion prediction, and 3D object detection from unstructured LIDAR data. Recently, selfsupervised learning has also shown promise for 3D scene flow estimation
[23, 24].The aforementioned 3D scene flow methods focus on accurately predicting pointtopoint correspondences. They often suffer from high computational costs, which is a key challenge for realtime deployment on a robotic platform.
IiB Occupancy Grid Maps
Occupancy grid maps (OGMs) are widely used to represent scene obstacle occupancy for robotics applications. Ondruska et al. [25] propose a deep tracking framework which incorporates a simple RNN to learn OGMtoOGM mappings. Ushani et al. [26]
formulate 2D BeV flow estimation as a similarity learning problem by transferring 3D OGMs into 2D embedding grids. A separate classifier learns the matched foreground cells between frames by using an expectation maximization algorithm. Later, Dequaire
et al. [27] extend the work of Ondruska et al. [25]by using a spatial transformer module and dilated gated recurrent units to account for observations from a moving platform. Wirges
et al. [28] propose a learned approach to determine a motion mask on an OGM but uses hand crafted input features such as mean intensity and height range of points falling within each cell, rather than raw point clouds.Estimation of the per cell motion state within an occupancy grid is a recent topic referred to as dynamic occupancy grid maps (DOGMa) estimation. Online versions typically model this state using particle filtering. Nuss et al. [29] propose an efficient implementation of DOGMa using a particle filtering scheme, which we adopt later on in the paper for comparison with our proposed method. Multiple methods have also been proposed to cluster and extract object level representations from a DOGMa for multiobject tracking [30, 31, 32, 33, 34]. Finally, some deep learning works build on the DOGMa representation for various tasks. For instance, Hoermann et al. [35] augment the DOGMa with a recurrent network trained by selfsupervised labeling to predict future states. Piewak et al. [36] build upon the Dynamic Occupancy Grid to do semantic segmentation of the DOGMa internal per cell state as static or dynamic.
While our method bears some similarities to the aforementioned works leveraging gridbased representation, we explore a new architecture bringing together both endtoend flow techniques and gridbased representations.
Iii Proposed System
We propose a method, PillarFlow, to learn to estimate 2D BeV flow by combining a Pillar Feature Network (PFN) [6] with a flow estimation network. The overview of the system is shown in Figure 2. First, two consecutive point cloud sweeps are aligned into the same coordinate frame: the original coordinates of the LIDAR sweep at are transformed to the coordinate frame of the LIDAR sweep at using the odometry information of the robot. Next, the two point clouds are encoded by PFN to build two BeV pseudoimages where each cell has a learned embedding based on points that fall inside it. Then the two pseudo images are fed to a flow network to estimate the dense 2D in the BeV space.
Iiia 3D Point Cloud to 2D BeV Embedding
In our system, we use a PFN to extract 2D BeV embeddings from 3D point clouds. First, a voxelization step is applied to the point cloud by discretizing the  plane, thus creating a set of “pillars” (grid cells) in birdseyeview. The voxelized pointcloud is structured as a
shaped tensor where
is the number of point descriptors, is the number of pillars, and is the number of points per pillar. We use , where the first four values denote coordinates and reflectance . The next five values are the distances to the arithmetic mean of all points in a pillar and the offset from the pillar center. Next, this input tensor is processed by a simplified version of PointNet [37] to get a feature map of shape . We further compress the feature map by a max operation over the last dimension, resulting in a encoded feature map with a dimensional feature embedding for each pillar. Finally, the encoded features are scattered back to original pillar locations to create a pseudoimage tensor of shape , where and indicate the height and width of the pseudoimage.IiiB 2D BeV Flow Network
To accurately associate the embeddings, i.e., pillar features between 2D BeV grids, we conduct a 2D BeV flow estimation. Based on the PWCNet model [11], we adjust architecture parameters such as receptive field and correlation layer parameters to account for the maximum relative motion that would be expected to be encountered between consecutive LIDAR sweeps (given the time delta between frames, grid resolution, and typical vehicle speeds).
As depicted in Figure 2
, the pillar features are further encoded via a feature pyramid network. A cost volume layer is then used to estimate the flow, where the matching cost is defined as the correlation between the two feature maps. Finally, a context network is applied to exploit contextual information for additional refinement. The context network is a feedforward CNN based on dilated convolutions, along with batch normalization and ReLU .
IiiC Training
We use annotated object tracks to generate 2D BeV flow ground truth. For each object, we estimate the instantaneous velocity from the difference in object positions divided by the elapsed time between consecutive frames. Our assumption is that only labeled dynamic objects can have a valid velocity, and all nonlabeled obstacles and background should have zero velocity. Note that this assumption might be violated in practice and does not provide direct supervision for potential outofontology moving objects (static objects have zero flow). Nonetheless, our experiments show that this provides enough supervision to learn an objectagnostic, dense (pillarlevel) optical flow network on BeV grids that generalizes well. Another exciting possibility we leave for future work would be to extend selfsupervised approaches like [24].
Let denote the flow field at the pyramid level predicted by the network with learnable parameters , and the corresponding ground truth. We use the common multiscale training loss from [38] and [11]:
(1) 
where is the L2 norm and the weights in the training loss are set by following the setting from Dequaire et al. [27].
Iv Experimental Results
We conduct a thorough evaluation and analysis at the 2D BeV flow estimation, and at the systemlevel by integrating our approach into different tracking systems.
Iva Datasets
To evaluate the performance of the proposed method, we conduct experiments on two datasets. First, we use nuScenes [39], a public largescale dataset for autonomous driving development, containing multiple sensors data such as LIDAR, radar, and cameras. The dataset includes 850 scenes for training and 150 scenes for validation, with fully annotated detection and tracking labels. Second, we use an inhouse dataset, TRIcuboid, collected by our fleet of autonomous cars equipped with LIDAR, radar, and several cameras. The dataset includes 194 scenes for training and 40 scenes for validation with fully annotated 3D bounding boxes for a variety of object categories.
IvB Implementation details
We limit the range of the point cloud to meters in both and directions with a grid resolution of meter per cell. During training, we use the Adam optimizer with exponentially decayed learning rate starting from and then reduce it by a factor of at each of total iterations. We train for 2M iterations on nuScenes, and 4M iterations on TRIcuboid dataset. For data augmentation, we apply (egocentric) random rotation and random bounding box scaling to the LIDAR sweeps. We perform simple ground plane removal to filter out points lying on a ground plane obtained with the RANSAC algorithm.
IvC 2D BeV Flow Estimation
To evaluate the performance of our PillarFlow model, we compare it to two baselines for BeV flow estimation.
Iterative Closest Point (ICP): ICP [40] outputs a transformation associating two point clouds by using an SVDbased pointtopoint algorithm. Here we select the 3D points within the clusters, and then apply ICP to obtain the transform between the two point clouds. The flow is inferred by the difference between the corresponding coordinates after the transformation.
Binary OGM
: Binary OGM binarizes the occupancy grid map from LIDAR sweeps. Instead of pillar features, we use the binary OGMs as the inputs of the adapted onechannel PWCNet, which learns to estimate the 2D flow in BeV grids.
Dataset  nuScenes  
[font=]Method Metric  RMSE (dynamic)  RMSE (static)  RMSE (average)  AAE 
ICP + Det.  2.818  NA  2.818  0.216 
Binary OGM  1.316  0.113  0.247  0.091 
Ours  1.127  1.110  0.207  0.108 
Dataset  TRIcuboid  
Method Metric  RMSE (dynamic)  RMSE (static)  RMSE (average)  AAE 
ICP + Det.  0.757  NA  0.757  0.681 
Binary OGM  0.409  0.036  0.081  0.098 
Ours  0.288  0.027  0.029  0.087 
Dataset  nuScenes  TRIcuboid  
Flow Network Used  RMSE  AAE  RMSE  AAE 
SpyNet  0.327  0.122  1.181  0.146 
FlowNet2  0.320  0.072  0.093  0.100 
PWCNet* (Ours)  0.207  0.087  0.029  0.087 
Ground Plane Config  RMSE  AAE  RMSE  AAE 
Ours w/ ground  0.246  0.079  0.069  0.092 
Ours w/o ground  0.207  0.087  0.029  0.087 
Metrics: We use root mean square error (RMSE) in m/s and average angular error (AAE) in radians as our metrics to measure flow error in 2D BeV grids.
Results: Table I shows the quantitative results comparing PillarFlow to the baselines, where the dynamic objects indicate movable objects such as vehicle, pedestrian, bicycle, truck, etc, and the static ones are immovable like building, pole, vegetation, etc. Our proposed method achieves an average error of in nuScenes and in TRIcuboid. Compared to the baselines, our method can better handle the flow produced by dynamic objects. Figure 3 shows qualitative results on the TRIcuboid dataset, confirming that the 2D BeV flows are accurately estimated. We observe that our method can deal better with empty cells and dynamic objects observed as stationary for which we predict zero flow while the baselines display significant noise on those objects.
To better analyze different components of our proposed approach, we compare against alternative flow network architectures, i.e., SpyNet [41] and FlowNet2 [38]. The comparison results are reported in Table II. The result indicates that PWCNet (the adaptation is mentioned in Section III) is the most compatible with the proposed system, showing a better capability of handling not only dynamic objects but also static ones. We also conduct another analysis of the impact of ground point removal, which yields an improvement of only RMSE on both nuScenes and TRIcuboid datasets. This also implies that the proposed method can perform properly even without filtering the ground plane.
Dataset  nuScenes  
[innerwidth=4cm]MethodMetric  AMOTA  AMOTP (m)  MOTA  MOTP (m) 
StanfordIPRLTRI [9]  56.1  80.0  48.3  37.9 
StanfordIPRLTRI + Binary OGM  56.5  79.6  48.6  38.2 
StanfordIPRLTRI + Ours  56.6  79.7  49.2  38.2 
IvD Tracking Evaluation on nuScenes
In this section, we provide an evaluation of our proposed 2D BeV flow estimation approach integrated in a stateoftheart tracking framework on nuScenes tracking dataset. We adopt the tracking pipeline of the winner entry to nuScenes Tracking Challenge at NeurIPS 2019 [42], StanfordIPRLTRI [9] as our baseline. The baseline takes 3D object detection results as measurement source and parameterizes them into a 7 dimension observation state: . To integrate our proposed 2D BeV flow estimation, we extend the observation state with 2D object velocity, . The object level velocity is approximated by the mean flow vector of all the BeV grids within the detected bounding box boundary. Under independent assumption between object detections and velocity estimations, our new observation model and covariance matrix are given as follows:
(2) 
(3) 
where R is the original noise covariance and is the noise covariance matrix for velocity. Following the suggestion by Chiu et al. [9], we estimate from the training set.
We report the tracking performance on the nuScenes tracking validation scenes in Table III
. We use the same evaluation metrics suggested by the nuScenes Tracking Challenge
[42], including average multiobject tracking accuracy (AMOTA), average multiobject tracking precision (AMOTP), multiobject tracking accuracy (MOTA), and multiobject tracking precision (MOTP). We compare our proposed approach to the baseline that did not use velocity in observation. We also compare to different velocity estimations provided by other BeV flow baselines (discussed in the Section IVB) in the same tracking algorithm. The results indicate that our estimated velocity is able to further improve the stateoftheart tracking performance working in parallel with 3D object detections. Compared with the velocity generated by binary OGM, our approach achieves better tracking stability (AMOTA) with less degradation in positional precision (AMOTP), which is a commonly seen tradeoff in multiobject tracking [42]. The better tradeoff also reflects the accuracy and robustness of our proposed 2D BeV flow estimation method.V Systematic Analysis on Autonomous Vehicle
In this section, we provide a fullsystem integration of a classagnostic tracking system and further discuss the advantage of our proposed method that is not fully captured in the previous metrics. We integrate the proposed method into an inhouse autonomous platform, and quantitatively evaluate it on 43 different 10 second snippet scenes collected from the Odaiba area of Tokyo, Japan, and Ann Arbor, Michigan, US. These logs have been fully annotated with bounding cuboids for ontology based objects. In this section, we briefly describe our tracking pipeline, and analyze how the proposed method improves the tracking performance of the generic objects.
Va Object Tracker Pipeline
To detect generic object clusters, we use a 2.5D terrain height map to filter out nonobject LIDAR points, by eliminating those falling outside the range of [m, m] above the corresponding ground cell height. Then, we collapse the remaining points to a 2D occupancy grid and utilize a connectedcomponents clustering algorithm to cluster the remaining points into classagnostic objects. We use a dilated object mask derived from a LIDAR object detector to enforce constraints on the labeling to ensure that the clusters are correctly segmented. Such a cluster representation can achieve a high object recall and ensure that all measurements are accounted for, which is important for safety, especially if a detection is not triggered.
The observation state of an object track is represented as . We apply a nearestneighbor association of existing tracks to object measurements by computing the Mahalanobis distance between the track’s predicted position and LIDAR object cluster’s position. A gating threshold on the nearest result is used to either construct an association, if close enough, or create a new track otherwise. To estimate the state of each object track, we use a fixedlag smoothing approach built upon a factor graph representation [43]. Binary factors between consecutive object state nodes encode a constant acceleration motion model. These factors attempt to minimize the difference in acceleration between consecutive states, with the error residuals weighed according to a predetermined square root information matrix. If a measurement (LIDAR object cluster) is associated to a track, a new state node is created and linked with the existing graph using this motion model prior. A unary prior factor is added to the new state node to represent the measurement, minimizing the difference between the measurement’s position and the corresponding state node’s position.
Metrics  Mean Track Velocity Error in m/s  95% Largest Track Velocity Error in m/s  
[innerwidth=4cm]CategoryMethods  Baseline  Baseline + DOGMa  Baseline + PF  Baseline  Baseline + DOGMa  Baseline + PF 
Static Objects  0.839  0.848  0.480  3.993  3.803  2.322 
Pedestrian & Cyclist  0.772  0.523  0.641  3.411  1.621  1.446 
Objects observed stationary  0.861  0.512  0.059  3.826  1.796  0.151 
Slow Moving Objects m/s  0.566  0.570  0.666  2.117  1.709  1.560 
Fast Moving Objects m/s  2.396  2.371  2.036  15.188  11.490  7.468 
VB Experimental Settings
We augment the tracker described above by adding an additional measurement prior, object velocity measurement, estimated from our proposed method (PillarFlow). We compare our estimated velocity with another nonlearning based velocity estimation methods widely adapted in realworld system, DOGMa [29]. Integration details adapting the two methods are given below:
PillarFlow
(PF): We aggregate the 2D BeV flows as estimated by our method to compute a single mean velocity and covariance per object cluster. This is simply obtained by random sampling the set of 2D BeV flows from the cells occupied by the object cluster.
DOGMa: DOGMa models the dynamic state estimation of grid cells as a random finite set (RFS) problem. We implement the DOGMa approach [29]
for comparison. Similarly, we aggregate the motion vectors to a mean per cluster by sampling, and weighing each sample based on the occupancy probability of the cell. Moreover, a weighted aggregation of the DOGMa’s preexisting velocity covariance matrix per occupied cell is applied to propagate the uncertainty of all cells. Note that this filtered approach can contain more information than just a two frame instantaneous estimate.
To compare the approaches, we evaluate the track versus ground truth object velocity error, instead of the typical tracking metrics such as MOTA and MOTP, false positive alarm rate, or track ID switch count. The reason is that generic objects are unlabeled and thus wouldn’t be reflected in the typical metrics. To associate a track with a ground truth annotated object for evaluation purposes, we require a significant overlap between the track’s most recent associated object cluster’s convex hull and a ground truth bounding box at the object cluster’s timestamp. Due to lack of bounding boxes for generic objects such as guardrails or trees, we assume that tracks that fail the association represent a generic LIDAR object with zero velocity.
VC Analysis and Discussion
The quantitative results in Table IV show strong enhancements to tracking performance using our proposed 2D BeV flow estimation as a prior. Overall, mean and worst case performance are improved across most object class types. In particular, the proposed method improves significantly in stationary objects (e.g., parked cars, standing pedestrians, and objects excluding static background).
We observe that the prior velocity information provides faster initialization and convergence of the track’s state estimate and allow for better data association, thereby creating more consistent tracks.
We provide qualitative results of the tracker integration in out autonomous platform, and an outofontology tracker observation in Figures 4 and 5 respectively.
Methods  ICP + Det.  DOGMa  PillarFlow 
mean (ms)  20  45  23 
variance (ms)  5  1  0.1 
VD Runtime Performance
Table V summarizes the runtime performance onboard our autonomous test platform. Our proposed model can achieve approximately 40Hz when run via TensorRT with halfprecision floating point mode on a single Nvidia Quadro RTX 6000 GPU. This demonstrates that the proposed method is feasible for realtime accurate velocity estimation in realworld applications.
Vi Conclusion
We propose PillarFlow, a deep network for endtoend dense motion estimation on 2D BeV grids. Experimental results show that PillarFlow improves the performance of dynamic object tracking on two datasets. Additionally, we demonstrate that PillarFlow delivers substantial improvements for realtime generic obstacle tracking onboard a realworld autonomous car. Nonetheless, we notice occasional incorrect flow predictions due to (dis)occlusions or for objects with few points. Interesting future work includes accumulating more temporal context or separately estimating occlusions and then augmenting the network input [44].
References
 [1] C. Badue, R. Guidolini, R. V. Carneiro, P. Azevedo, V. B. Cardoso, et al., “Selfdriving cars: A survey,” arXiv:1901.04407, 2019.
 [2] X. Chen, H. Ma, J. Wan, B. Li, and T. Xia, “Multiview 3D object detection network for autonomous driving,” in CVPR, 2017.
 [3] J. Ku, M. Mozifian, J. Lee, A. Harakeh, and S. L. Waslander, “Joint 3D proposal generation and object detection from view aggregation,” in IROS, 2018.
 [4] B. Yang, W. Luo, and R. Urtasun, “Pixor: Realtime 3D object detection from point clouds,” in CVPR, 2018.
 [5] Y. Zhou and O. Tuzel, “VoxelNet: Endtoend learning for point cloud based 3D object detection,” in CVPR, 2018.
 [6] A. H. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “PointPillars: Fast encoders for object detection from point clouds,” in CVPR, 2019.
 [7] D. Held, J. Levinson, S. Thrun, and S. Savarese, “Combining 3D shape, color, and motion for robust anytime tracking.” in RSS, 2014.
 [8] X. Weng and K. Kitani, “A baseline for 3D multiobject tracking,” arXiv:1907.03961, 2019.
 [9] H.k. Chiu, A. Prioletti, J. Li, and J. Bohg, “Probabilistic 3D multiobject tracking for autonomous driving,” arXiv:2001.05673, 2020.
 [10] A. Buehler, A. Gaidon, A. Cramariuc, R. Ambrus, G. Rosman, and W. Burgard, “Driving through ghosts: Behavioral cloning with false positives,” in IROS, 2020.
 [11] D. Sun, X. Yang, M.Y. Liu, and J. Kautz, “PWCNet: CNNs for optical flow using pyramid, warping, and cost volume,” in CVPR, 2018.
 [12] S. Baker, D. Scharstein, J. Lewis, S. Roth, M. J. Black, and R. Szeliski, “A database and evaluation methodology for optical flow,” IJCV, vol. 92, no. 1, pp. 1–31, Nov. 2011.
 [13] A. Dewan, T. Caselitz, G. D. Tipaldi, and W. Burgard, “Rigid scene flow for 3D lidar scans,” in IROS, 2016.
 [14] A. K. Ushani, R. W. Wolcott, J. M. Walls, and R. M. Eustice, “A learning approach for realtime temporal scene flow estimation from LIDAR data,” in ICRA, 2017.
 [15] V. Vaquero, A. Sanfeliu, and F. MorenoNoguer, “Deep Lidar CNN to understand the dynamics of moving vehicles,” in ICRA, 2018.
 [16] S. A. Baur, F. Moosmann, S. Wirges, and C. B. Rist, “Realtime 3D LiDAR flow for autonomous vehicles,” in IV, 2019.
 [17] X. Gu, Y. Wang, C. Wu, Y. J. Lee, and P. Wang, “HPLFlowNet: Hierarchical Permutohedral Lattice FlowNet for scene flow estimation on largescale point clouds,” in CVPR, 2019.

[18]
S. Wang, S. Suo, W.C. Ma, A. Pokrovsky, and R. Urtasun, “Deep parametric continuous convolutional neural networks,” in
CVPR, 2018.  [19] X. Liu, C. R. Qi, and L. J. Guibas, “FlowNet3D: Learning scene flow in 3D point clouds,” in CVPR, 2019.
 [20] C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “PointNet++: Deep hierarchical feature learning on point sets in a metric space,” in NeurIPS, 2017.
 [21] Z. Wang, S. Li, H. HowardJenkins, V. A. Prisacariu, and M. Chen, “FlowNet3D++: Geometric losses for deep scene flow estimation,” arXiv:1912.01438, 2019.
 [22] A. Behl, D. Paschalidou, S. Donné, and A. Geiger, “PointFlowNet: Learning representations for rigid motion estimation from point clouds,” in CVPR, 2019.
 [23] H. Mittal, B. Okorn, and D. Held, “Just go with the flow: Selfsupervised scene flow estimation,” arXiv:1912.00497, 2019.
 [24] W. Wu, Z. Wang, Z. Li, W. Liu, and L. Fuxin, “PointPWCNet: A coarsetofine network for supervised and selfsupervised scene flow estimation on 3D point clouds,” arXiv:1911.12408, 2019.

[25]
P. Ondruska and I. Posner, “Deep tracking: Seeing beyond seeing using recurrent neural networks,” in
AAAI, 2016.  [26] A. K. Ushani and R. M. Eustice, “Feature learning for scene flow estimation from LIDAR,” in CoRL, 2018.
 [27] J. Dequaire, P. Ondrúška, D. Rao, D. Wang, and I. Posner, “Deep tracking in the wild: Endtoend tracking using recurrent neural networks,” IJRR, vol. 37, no. 45, pp. 492–512, 2018.
 [28] S. Wirges, J. Gräter, Q. Zhang, and C. Stiller, “Selfsupervised flow estimation using geometric regularization with applications to camera image and grid map sequences,” arXiv:1904.12599, Apr. 2019.
 [29] D. Nuss, S. Reuter, M. Thom, T. Yuan, G. Krehl, M. Maile, A. Gern, and K. Dietmayer, “A random finite set approach for dynamic occupancy grid maps with realtime application,” IJRR, vol. 37, no. 8, pp. 841–866, July 2018.
 [30] R. G. Danescu, “Obstacle detection using dynamic particlebased occupancy grids,” in DICTA, 2011.
 [31] S. Steyer, G. Tanzmeister, and D. Wollherr, “Object tracking based on evidential dynamic occupancy grids in urban environments,” in IV, 2017.
 [32] A. Vatavu, N. Rexin, S. Appel, T. Berling, S. Govindachar, et al., “Environment estimation with dynamic grid maps and selflocalizing tracklets,” in ITSC, 2018.
 [33] F. Gies, A. Danzer, and K. Dietmayer, “Environment perception framework fusing multiobject tracking, dynamic occupancy grid maps and digital maps,” in ITSC, 2018.
 [34] S. Steyer, C. Lenk, D. Kellner, G. Tanzmeister, and D. Wollherr, “Gridbased object tracking with nonlinear dynamic state and shape estimation,” TITS, vol. 21, no. 7, pp. 2874–2893, July 2019.
 [35] S. Hoermann, M. Bach, and K. Dietmayer, “Dynamic occupancy grid prediction for urban autonomous driving: A deep learning approach with fully automatic labeling,” in ICRA, 2018.
 [36] F. Piewak, T. Rehfeld, M. Weber, and J. M. Zollner, “Fully convolutional neural networks for dynamic object detection in grid maps,” in IV, 2017.
 [37] R. Q. Charles, H. Su, M. Kaichun, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in CVPR, 2017.
 [38] E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovitskiy, and T. Brox, “FlowNet 2.0: Evolution of optical flow estimation with deep networks,” in CVPR, 2017.
 [39] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, “nuScenes: A multimodal dataset for autonomous driving,” arXiv:1903.11027, 2019.
 [40] P. J. Besl and N. D. McKay, “Method for registration of 3D shapes,” in TPAMI, vol. 14, no. 2, Feb. 1992.
 [41] A. Ranjan and M. J. Black, “Optical flow estimation using a spatial pyramid network,” in CVPR, 2017.
 [42] “nuScenes tracking task benchmark.” [Online]. Available: https://www.nuscenes.org/tracking/
 [43] F. Dellaert, M. Kaess, et al., “Factor graphs for robot perception,” Foundations and Trends in Robotics, vol. 6, no. 12, pp. 1–139, 2017.
 [44] E. Ilg, T. Saikia, M. Keuper, and T. Brox, “Occlusions, motion and depth boundaries with a generic network for disparity, optical flow or scene flow estimation,” in ECCV, 2018.