Research On Pedestrian Detection Algorithm Based On YOLO

With the progress and development of computer science and technology,a lot of manpower completed work and tasks,can now be completed by computer vision.At present,computer vision is a very important research direction and research hotspot.The research in this field involves many fields and disciplines,such as computer image video processing,machine learning,linear algebra,etc.One of the ultimate goals of the research is to be able to simulate human visual recognition to complete various identification tasks.Pedestrian detection,as a popular research direction of computer vision,mainly includes pedestrian tracking,abnormal behavior monitoring and pedestrian recognition.A good pedestrian detection algorithm can provide strong theoretical support and technical support for the research of pedestrian recognition.In recent years,deep learning technology and machine learning technology have also made a breakthrough in the field of visual inspection.On this basis,this paper will study,summarize and make use of some excellent achievements of scholars at home and abroad.The paper optimize the prior box selection method in the YOLOv3 target detection network and add deformable convolution layer and deformable pooling layer.It uses semantic model to assist pedestrian detection to improve the accuracy of pedestrian detection.The method of lightweight model and model pruning is used to modify the network infrastructure and to improve the real-time performance of the algorithm.The main contents of this paper are as follows:Firstly,in the view of the fact that pedestrian detection is easy to be affected by illumination,background interference,occlusion and other factors,a deformable convolution layer is added to the YOLOv3 target detection network.The region of interest is input to the pooling layer to pool the mean value of the deformable position sensitive.The classification and target position refinement are carried out.A method to pool the location sensitive region of interest alignment is proposed.The learning of target features is also enhanced.Secondly,it is found that YOLO did not fully consider the classification and confidence of pedestrians in the selection of the prior box.IOU(Intersection-over-Union,IOU)is chosen as the basis and the only standard of the prior box evaluation.Therefore,this paper mainly optimizes the prior box selection.When selecting of priori box,the confidence between pedestrian classification of IOU is considered more comprehensively.The current pedestrian detection algorithm,the target image of YOLOv3 is often used to overcome the shortcoming of high rate of missing detection,existing in the image detection network.Because of the certain relevance between the pedestrian distribution in the target image network and its semantic counting attribute learning,this paper proposes an image depth learning method,which integrates the pedestrian detection semantics and the semantic counting of pedestrian detection methods.At the same time,by detecting the distribution of pedestrians and pedestrians in the image and the attributes of their semantic counting methods,the method of learning the semantic attributes of pedestrians in the image is used to assist the detection of pedestrians in the image.It can suppress the influence and interference of the semantic attributes of pedestrians in the image on pedestrians and can improve the detection accuracy.Aiming at improving the accuracy of pedestrian counting by depth learning in the scene of target image video detection,this paper makes in-depth experimental and theoretical analysis.Combining YOLO and DeepSort is put forward to get the pedestrian counting by depth learning.Thirdly,real-timely processing of pedestrian detection and the complexity of the model is often needed.Shufflenet is used to replace the feature extraction network in Dark Net53 network and prune the model.The optimized YOLOv3-W-D network can significantly reduce the size of the model and the computational complexity of the model without reducing the accuracy.The real-time performance of the algorithm is also effectively improved.By the application of these methods,the precision and real-time performance of the improved YOLOv3 network in pedestrian detection,are improved obviously.It shows the effectiveness of the research method in this paper.