Lights as Points: Learning to Look at Vehicle Substructures With Anchor-Free Object Detection

Vehicle detection is a paramount task for safe autonomous driving, as the ego-vehicle must localize other surrounding vehicles for safe navigation. Unlike other traffic agents, vehicles have necessary substructural components, such as the headlights and tail lights, which can provide important cues about a vehicle's future trajectory. However, previous object detection methods still treat vehicles as a single entity, ignoring these safety-critical vehicle substructures. Our research addresses the detection of substructural components of vehicles in conjunction with the detection of the vehicles themselves. Emphasizing the integral detection of cars and their substructures, our objective is to establish a coherent representation of the vehicle as an entity. Inspired by the CenterNet approach for human pose estimation, our model predicts object centers and subsequently regresses to bounding boxes and key points for the object. We evaluate multiple model configurations to regress to vehicle substructures on the ApolloCar3D dataset and achieve an average precision of 0.782 for the threshold of 0.5 using the direct regression approach.