Abstract:
As a growing shift in autonomous driving, in addition to 3-dimensional perception of the surrounding area, the dynamics of the objects are important to result in better driving performance with low collisions. Some existing works have considered optical flow for autonomous driving and have tracked the trajectory of vehicles using a bird’s-eye view. These works lack focus on regions of interest (RoI) as they track objects without prioritization, can suffer from object occlusions, and suffer from situational non-awareness. Thus, in order to cater to this research gap, in this work, we propose to use the driver’s perspective to first obtain the RoI by identifying moving objects - pedestrians and vehicles - using a novel divide-and-conquer algorithm to generate multiple level 2D bounding boxes using a segmented image, tackling overlapping segments using pixel matching. Next, for the determined RoI, we generated the optical flow by considering a weighted energy functional to prioritize RoI compared to other areas, ensuring data fidelity and smoothness in the interested areas. Finally, two segmented images are utilized for motion vector generation, where we perform the augmented template normalized cross-correlation in the neighborhood of RoI to determine the motion of objects, which mitigates the effect of object changing (using augmentation), false positives (using neighborhood search), and partial template (using template splitting) problems. After template matching, the centroids of the previous object and new object are utilized to derive the motion vectors, which represent the relative motion of objects with respect to the driving vehicle. We obtain a dataset from the CARLA simulator and the KITTI dataset to evaluate the performance of the proposed technique. The results show that motion vectors resemble the real relative velocities of the objects (error < 7.5%), and the proposed RoI determination (error < 6.8%) and optical flow finding (error < 4.5%) models are effective.
