Detection And Application Of Pedestrian Self-occlusion Based On Deep Learning

In recent years,pedestrian detection has become a research hotspot in the field of computer vision.Pedestrian detection can be defined as detecting the presence of pedestrians in an input picture or video frame.This technology provides strong support for autonomous driving,video surveillance,and human behavior analysis.Recently,it has also been used in emerging fields such as victim rescue and aerial photography,and has a wide range of applications.The study in this thesis is mainly based on Faster R-CNN(Faster Region-Convolutional Neural Networks)target detection algorithm,and we redesign its network structure to improve the detection performance in pedestrians' partial occlusion problems.Structures of this theis are as follows:The dissertation first summarizes the current research difficulties of pedestrian detection and the current research in and out of China,and investigates and combs traditional pedestrian detection methods and target detection algorithms based on deep learning.Among them,the most classic of the traditional methods is the algorithm model based on HOG + SVM(Histogram of Oriented Gradients+Support Vector Machine).The target detection algorithms based on deep learning are mainly divided into three categories: SSD(Single Shot Multi Box Detector),R-CNN series and YOLO(You Only Look Once)series.Secondly,by comprehensively considering the detection performance and speed of the algorithm,a pedestrian detection algorithm based on the Faster R-CNN target detection framework was trained using the self-built occluded pedestrian dataset,then we use the test set to experiment on the model based on HOG+SVM and the trained Faster R-CNN model,and the experimental results are compared.The experimental results show that the detection effect of Faster R-CNN model is obviously better than traditional detection methods,and the features learned by deep convolutional neural network are more robust.Finally,the network structure of Faster R-CNN model is modified for occluded pedestrian detection.(1)To improve the feature network of Faster R-CNN,we select Res Net(Residual Network)in combination with SENet(Squeeze-and-Excitation Networks)for feature extraction;(2)We redesign the aspect ratio of anchor in RPN(Region Proposal Network)network to make it easier to match pedestrians;(3)By considering that pedestrians are Prone to self-occlusion and environmental occlusion problems,denser pedestrian data is added to the training set,and repulsion Loss is used to improve the model's ability to detect occluded pedestrians.The final results show that the improved Faster R-CNN model upgrades the detection performance with a small increase in detection time.