Research On Lightweight And Occlusion Pedestrian Detection Methods Based On Deep Learning

In recent years,with the rapid development of deep learning technology,pedestrian detection research based on deep learning has received much attention.Pedestrian detection is a sub-task of object detection for pedestrian class,which has been a difficult and hot research topic in the field of computer vision.Pedestrian detection is not only of high scientific value,but also has high application value in the fields of advanced assisted driving,intelligent surveillance and intelligent robotics.Although deep learning-based object detection has been well developed,the following problems still exist for pedestrian detection tasks: in practical applications,the computing capability and power consumption of the platform will be limited,and how to ensure the balance between detection accuracy and detection speed of pedestrian detection models has become an urgent problem to be solved.In addition,due to the uncertainty of occlusion patterns and the commonality of intra-class occlusion in dense crowds,there are still more missed detections in the current pedestrian detection models.This paper addresses the problem of detection speed of pedestrian detection models in hardware-constrained environments and the effective detection in occlusion scenarios,and makes several improvements based on existing ideas of object detection networks to make them have superior pedestrian detection performance.This paper mainly covers the following:(1)For the problem of slow pedestrian detection under hardware limitation,this paper proposes a lightweight pedestrian detection model based on YOLOv4.In this model,the lightweight backbone network Ghost Net is used to replace the CSPDark Net53 using the cheap Ghost module to produce more feature maps,which can effectively reduce the size of the model as well as the computational effort;the Focal-EIo U loss function is used to more directly use the edge length as the penalty term,which improves the model training effect by focusing more on high-quality pedestrian instances during training.The model neck structure PANet was pruned to further balance the model size between the backbone network and the neck structure;the model was deployed on the Jetson Nano embedded platform and accelerated using Tensor RT technology.The experimental results show that the present model significantly improves the inference speed with a slight improvement in the miss detection rate compared to YOLOv4.(2)For the problem of high miss detection rate in occlusion scenes.The Multi-Attention Context Network(MACNet)is proposed.The network improves pedestrian feature extraction by adding multi-attention context module to make the detector obtain richer contextual information and use its attention mechanism to learn different occlusion patterns;establishes adaptive mutual supervision mechanism by adding trainable parameters combining global context module and multi-attention context module;we propose an occlusion perception force loss function to solve the pedestrian intra-class occlusion problem by borrowing the design idea of focal repulsion loss;ignores unreasonable samples and too small samples in the training process to reduce the interference of such samples to the detector In the training process,we ignore unreasonable samples and too small samples to reduce the interference of such samples to the training of the detector.The experimental results show that the method reduces the missed detection rate in different scenarios,and the improvement of the missed detection rate in severe occlusion is more obvious.