Research On Pedestrain Detection Technology Based On Deep Learning

Pedestrian detection is an important part of intelligent video analysis.Computer vision technology is used to determine whether there is a pedestrian in the image or video sequence and give accurate positioning.Due to the influence of scale and occlusion between pedestrians,pedestrian detection methods have some problems,such as poor feature extraction effect,low positioning accuracy and slow detection speed.Therefore,pedestrian detection is widely concerned by many researchers.This paper focuses on the pedestrian detection method based on deep learning,and researches on small pedestrian detection,dense pedestrian detection and other aspects,as follows:In the aspect of pedestrian detection based on convolutional neural network,three pedestrian detection methods based on Faster R-CNN,Cascade R-CNN and RetinaNet are studied and implemented,and evaluated on Caltech dataset.The experimental results show that the one-stage RetinaNet method is superior to Faster R-CNN and Cascade R-CNN in AP50,Recall and pedestrian detection efficiency,but its too dense anchors make mMR and F1 score underperformance.In the aspect of small pedestrian detection,an improved small pedestrian feature extraction method is proposed.First of all,by adjusting the distance measurement parameters of Kmeans clustering,we get the anchor box which has a higher fit with the groundtruth and solve the problem of high complexity in the training process of convolution neural network feature extraction and boundary box regression training.Secondly,a more powerful dense module is introduced into the backbone convolutional neural network used in the YOLOv3 framework to enhance the ability of the algorithm to extract the characteristics of small pedestrians and solve the problem that the characteristics of small pedestrians are not easy to distinguish from the complex background.Then,the test experiment is carried out on Caltech dataset,and miss rate is reduced from 12.79% to 11.29%.The results show that the improved method can reduce the miss rate of small pedestrian target.Finally,the average accuracy of the VOC2007 Person subset,VOC2012 Person subset and UAVDT dataset is 84.1%,85.4% and 42.03%,respectively.The experimental results show that the improved method has good robustness and generalization capability.In the aspect of dense pedestrian detection,an improved precise location regression method is proposed.First of all,two loss functions,DIoU and CIoU,are used to replace the MSE loss function used by the original YOLOv3 in the process of bounding box regression,so as to accelerate the convergence speed,improve the location accuracy and solve the problem of poor detection performance caused by dense pedestrian location offset.Secondly,in the post-processing stage of detection,the non maximum suppression based on DIoU is used to replace the traditional non maximum suppression,which can keep more correct detection boxes without increasing too many false-positive boxes,and solve the problem of reducing the recall rate caused by the traditional non maximum suppression operation.Finally,the improved method is tested on the Crowd Human dataset,and the results are 72.77%,43.42% and 42.92% on AP50,AP75 and AP respectively.In the aspect of pedestrian detection software implementation,using C language and QT creator 5.14.1 development platform,the pedestrian detection based on deep learning is realized by programming,and the data collected in the real scene is used for verification test.The software includes video pedestrian detection,single image pedestrian detection,dataset performance evaluation and other functions.