Pedestrian Detection with Simplified Depth Prediction

Though RGB Cameras, Radar and LIDARs are popular sensors for intelligent vehicle systems, real-time joint inference on their sensory outputs remains challenging. Moreover, high-resolution LIDAR is expensive both in terms of cost and computation. This paper presents a deep learning-based pedestrian detection algorithm that takes both RGB image and lower-resolution LIDAR data and returns object detections in the image as 2-D bounding boxes, plus the distances of the detected objects. The proposed network is much less expensive but comparable in accuracy to previous deep networks that combine these sensors use image-like or voxel representations of LIDAR data to directly predict 3D positions and shapes. To train this network, a new dataset was created, containing register information from low-end camera a 16-layer LIDAR, and corresponding ground truth distance values generated by estimating the position of pedestrians from global navigation satellite system (GNSS) sensors and a fixed tower. The public release of this dataset is an additional contribution of this effort.