OpCNet: Endowing vehicles with perspective vision: Clairvoyance of occluded Pedestrian crossing in complex driving scenes

A major safety challenge for intelligent autonomous vehicle is the perception of sudden and unpredictable occluded pedestrian crossing (OPC). So far, the performance on heavily vision-obstructed pedestrians detection in complex driving scenes remains unsatisfied. Instead of focusing on the detections of unobstructed parts of pedestrians in previous works, our efforts have been made to explore solution from a novel path by restoration and completing pedestrian from occlusions. We proposed a self-supervised learning model termed OpCNet which consists of two main modules of instance segmentation and partial completion network. Our proposal demonstrated quality performance in amodal segmentation and making completion-inference on heavily occluded pedestrians in complex traffic scenes. This could effectively helps to improve the detection and behavior predictions of occluded pedestrians.