Research On Dense Pedestrian Detection Algorithm Based On YOLO

In today's society,image processing algorithms based on artificial intelligence and computer vision are changing day by day.Pedestrian detection using deep learning has become a current research hotspot.However,there is a certain degree of occlusion problems for pedestrian targets in dense scenes,which restricts the development of classic deep learning detection methods.For this reason,this paper studies intensive pedestrian detection based on deep learning,which effectively improves the performance of pedestrian detection in dense scenes.The work content of this article can be summarized as follows:First,in response to the missing detection problem caused by insufficient feature extraction capabilities of the original YOLOv3,an improved network model YOLO-VAS based on YOLOv3 is proposed.The main improvements include: 1)First,by improving the feature extraction network and introducing an efficient attention mechanism to enhance the network's ability to extract important information;2)adding an improved SPP module in the feature fusion stage to achieve the fusion of local features and global features.Enrich the expression ability of the final feature map;3)Based on the specific pedestrian data set,the anchor box is re-clustered using the K-means mean algorithm to obtain 9 new anchor boxes;4)Improve the bounding box regression loss function of the original network,reduce the missed detection rate and accelerate the convergence of the model;5)Improve the original candidate box screening algorithm NMS,which effectively reduces the pedestrian missed detection rate in dense scenes;6)Finally Through a series of data enhancement strategies,the diversity of the data set is enriched and the generalization ability of the model is improved.Experimental results show that compared with the original YOLOv3 network,the improved YOLO-VAS model can effectively improve the detection performance in dense pedestrian scenes.Second,on the basis of the simplified version of YOLOv3 model YOLOv3-tiny,the network structure is improved,and the lightweight model is designed.1)An improved cheap operation method with an efficient attention mechanism uses a few parameters to generate more important feature maps;2)Aiming at the problem of the original network's missed detection of small target pedestrians,by adding a predictive scale to improve The model's ability to detect small target pedestrians.The experimental results show that the improved network can further improve the detection accuracy under the premise of high real-time performance.