Research On Algorithm Of Dense Pedestrian Detection Based On YOLOv3

Pedestrian detection,as an important branch of target detection,has been widely used in behavior analysis,traffic monitoring,driverless technology and intelligent medical treatment.Deep learning detection based on convolutional neural network features has become the mainstream method in the field of detection technology due to its strong self-learning ability and self-adaptability,and has also made outstanding achievements in the field of pedestrian detection,this is especially true for sparse,large-scale pedestrian targets.The distance between the pedestrian and the camera affects the size of the target when the crowd scene is applied to the practical application.The small pedestrian targets are mostly captured,and they are more easily influenced by the environment noise and occlusion,the detection effect is obviously different from the large scale sparse target,so there is a higher demand for the model.This paper focuses on dense pedestrian detection based on Yolov3.(1)We mainly study the algorithm of object detection based on convolutional neural network,select the Yolov3 network based on regression as the basis of network model,analyze the core of YOLO algorithm and analyze its advantages and disadvantage s,through data processing,the network structure and the loss function are improved and the dense pedestrian detection model is built to optimize the detection performance on small target and multi-scale.(2)To improve the robustness of the model,the sample features are enriched by means of reinforcement before training,and the anchor frame is re-selected according to the characteristics of dense crowd.In view of the low detection accuracy and high miss rate caused by the poor representation of small targets,a deep feature extraction is carried out by FLPN auxiliary network with strong position feature pyramid structure,FPLN can get stronger position information by aggregating the information of the th ree feature maps and then bi-directionally creating feature information stream.In combination with the introduction of penalty bounding-box regression loss function,the offset of the prediction box is reduced by narrowing the distance between the center of the target box and the prediction box,and the accuracy of the box is improved.(3)The performance of the dense pedestrian detection model based on Yolov3 is evaluated by experiments on image and video data.The average accuracy of the small target is 92.25%,the average accuracy is improved by 4.25%,and the overall average accuracy is 94.65%,it's also real-time.Compared with the original Yolov3 model,the precision of Human Crowd is improved by 6.82% and the average miss rate is reduced by 2.32%.It is proved that the proposed dense pedestrian detection model can extract features in a deeper level,and has validity and generalizability.