Research On Crowded Pedestrian Detection Based On Deep Learning

As a famous research field of computer vision,pedestrian detection technology plays an important role in many fields such as video surveillance and automatic driving.With the development of deep learning,pedestrian detection accuracy is greatly improved,However,the pedestrian detection in crowded scenes is still difficult.,Firstly,the similarity between pedestrians is high,while current detection models focus on extracting overall features.This makes it difficult to distinguish between highly overlapped pedestrians.Secondly,there are limitations in the post-processing methods of prediction boxes.Such as Faster R-CNN,YOLOV3 detection models collect samples from feature maps to generate dense prediction boxes.Non-Maximum Suppression(NMS)is adopted for removing overlapped prediction boxes.However,it is very difficult to set an NMS threshold when this method is applied to crowded pedestrian scenarios.If the NMS threshold is too low,a large number of missed detections will be generated.If the NMS threshold is too high,a large number of false detections will be generated.However,the current common detection methods for crowded pedestrians are often improved for only one difficulty.Therefore,this thesis proposes an improved Mask R-CNN network to deal with the two difficulties at the same time,mainly doing the following three aspects:(1)Proposing SFPN module to strengthen the edge feature extraction ability of crowded pedestrian.(2)Combining head bounding box and body bounding box to construct a more refined pseudo semantic segmentation label.(3)Designing a rule that used to estimate the image pedestrian visibility according to the human keypoints,only the image with low pedestrian visibility will re-input to network to detect occluded pedestrians,which significantly reduced detection time.Experiments on Crowd Human and Wider Person pedestrian detection datasets in crowded scenarios verify the effectiveness of the improved Mask R-CNN network in this thesis,which is superior to other popular methods in terms of miss rate and accuracy.