DTA: Detect Them All for Safe and Reliable Autonomous Driving

Object detection is an essential component of autonomous driving systems, as it provides an initial perception of objects in the surroundings. These detections are used to track and model the behavior of traffic participants, thereby ensuring the safety of everyone by avoiding collisions. Therefore, for reliable autonomous driving, both common and uncommon objects must be detected. The existing approaches for object detection are constrained to predict common classes based on the dataset utilized for their training. Further, their predictions are unverified and unreliable, as these approaches are based on deep neural networks, which are not transparent. Moreover, the commonly used object detection evaluation metric, mAP, equally weighs the objects detections which are in front of ego vehicle and those which are far away. To address these issues, we propose an effective and intuitive multimodal approach that uses an object detector as a base to obtain initial predictions. It employs an intuitive ground plane estimation and uses it to filter LiDAR points with distance, overhang, and ground clearance thresholds. The resulting points are clustered and utilized to verify initial detections; the unmatched clusters are incorporated into final detections as objects with a default class. We also introduce <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$mAP_{c}$</tex>, an evaluation measure for object detection in traffic scenes that considers the critical objects. To evaluate the effectiveness of detection verification, we compare initial detections with verified detections using <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$mAP_{c}$</tex>, where it shows significant improvement.