Research On Pedestrian Detection Method Based On Residual Network And Multi-scale Training Strategy

Pedestrian detection based on image and video is the foundation and premise of pedestrian tracking,behavior analysis,gait analysis,pedestrian identity recognition,and also the most important research direction in the field of computer vision.With the development of the times,whether it is video surveillance,vehicle assisted driving or intelligent robot,people's demand for pedestrian detection is growing.In recent years,pedestrian detection algorithms based on deep learning have been proposed to solve the shortcomings of traditional manual methods,such as low accuracy and slow speed.The existing pedestrian detection algorithm based on deep learning takes up a lot of resources,and its accuracy and speed need to be further improved,which can't meet the needs of realtime monitoring.Based on the prototype of micro YOLO,this thesis studies the pedestrian detection based on YOLO algorithm,and improves it.SSE algorithm is used to cluster the candidate frames,and multi-scale training strategy is used to train on the mixed data set.The improved model takes up less storage space,further improves the accuracy and speed of detection,and enhances the generalization ability of the model.The specific work is as follows:(1)This thesis expounds the background and significance of pedestrian detection research,and studies the current situation of traditional pedestrian detection and deep learning pedestrian detection at home and abroad;introduces the related technologies of pedestrian detection,including pedestrian detection based on traditional methods and deep neural network;analyzes the micro YOLO algorithm in detail,and gives the initial experimental results on the INRIA standard data set.(2)A pedestrian detection model based on improved YOLO network structure is proposed.The yolov3 tiny network structure is improved in different ways.In order to speed up the detection,a fine-grained multi-scale fusion method is adopted.At the end of the original network structure,a feature map output is added to form the new network structure yolov3-tiny2.Secondly,in order to avoid the defect of information loss in the detection process,a volume layer and two residual modules are added to the backbone network of the original network model to form the network model yolov3-tiny3.Finally,the combination of the two is verified whether it will have a greater detection advantage,combine the above two methods to get yolov3-tiny23.Through the experimental comparison and analysis,the network improvement method proposed in this paper has improved in accuracy,detection speed and space resource occupation.(3)A pedestrian detection model based on multi-scale training of mixed dataset is proposed.The k-value of K-means clustering algorithm is optimized by SSE,and the classifier is trained on user-defined data set and standard data set Pascal VOC and INRIA through the optimization of network training parameters and multi-scale training strategy.The experiment shows that the multi-scale training method of mixed data sets can maintain the detection speed,and improve the accuracy,recall,and the average intersection ratio of the bounding box.