Dynamic Occlusion Perception Data Enhancement for Occluded Pedestrian Detection

Abstract

In the automatic driving scene, pedestrian occlusion detection has always been a difficulty. Due to the limited diversity of occlusion patterns in the training data, existing methods perform poorly when encountering various occlusion changes in complex scenes. To solve this problem, this paper proposes a dynamic occlusion-aware data augmentation method. Its core is to add real object occlusion for pedestrian instances in real time during the training process, and update the visible box annotation synchronously. The specific approach is to use the Coco dataset to build an occlusion database containing more than 5000 real object instances, and each instance is saved as an RGBA image to retain the shape and transparency information. Then, probability enhancement is performed on pedestrians in each training image with a probability of 50%. Finally, the transparency channel is used to complete image fusion. This method can be directly embedded into the existing pedestrian detection framework. Experimental results show that on CityPersons and Caltech datasets, DOA reduces the miss detection rate of the benchmark model on the severe occlusion subset by 3.2% and 2.7%, respectively, which proves that this method plays a key role in improving the occlusion robustness.