Autonomous Driving Perception

Autonomous driving perception turns sensor data into estimates of what is around a vehicle, where things are, and how they are moving, supporting safe navigation and interaction. Common tasks include 2D and 3D object detection, lane and road segmentation, tracking, depth estimation, and 3D occupancy; approaches use cameras, LiDAR, radar, event sensors, or sensor fusion, often with bird’s-eye-view representations and temporal models. Methods are commonly evaluated on datasets such as nuScenes, Waymo, KITTI, Cityscapes, and OpenLane, with increasing attention to night, adverse weather, long range, occlusion, and sensor failure. A central challenge is improving real-world reliability and safety without excessive compute: learned fusion and vision-language models can add context, but must be grounded in accurate geometry and tested beyond clean, familiar benchmark conditions.