Pedestrian Detection In Campus Real Scene Based On Deep Learning

Pedestrian detection has always been a hot issue in the field of target detection,and the combination of deep learning and pedestrian detection excavates the potential research and application value of pedestrian detection.This paper is based on deep learning to effectively detect the pedestrian in the campus scene.The traditional pedestrian detection algorithm can not meet the accuracy and real-time requirements of pedestrian detection,so this paper uses the YOLOv4 network model just released in 2020 to realize the real-time detection of pedestrians on campus,and makes the preliminary research preparation for the combination with campus video security surveillance in the future.The main research contents of this paper are as follows:(1)Based on the full study and comparison of several commonly used target detection algorithms based on deep learning,this paper chooses to use YOLOv4 network model to carry out campus pedestrian detection experiment.The network model trained by YOLOv4 not only shows high accuracy in pedestrian detection,but also can real-time detect pedestrians in campus video images,which can not be compared with traditional algorithms.(2)This paper makes a video image pedestrian data set which can be trained by YOLOv4 network model under the campus background.The pedestrian image material in the data set is captured by the high-definition camera of the mobile phone,which is shot in different places and different time periods in the school,and the shooting angle and distance are also different.It includes school pedestrians with different angles and different heights,and then the video image is intercepted at the rate of 30 frames per interval as the original material of the data set.And use the professional tagging software——Label Image,to make sample tags for pedestrians in the data set,and the tags are saved in a fixed format document,which contains the category and location information of the samples.(3)Aiming at the width-height ratio of pedestrian samples in the pedestrian data set in this paper,we no longer use the original K-means clustering algorithm in YOLO network,but use the improved clustering algorithm K-means++ to re-cluster the data set and select a prior frame size that is more in line with the sample label size.The experimental results show that K-means++ algorithm is more helpful to improve the accuracy of network detection than K-means algorithm.(4)In order to train the YOLOv4 network model on the general GPU,the original network model is properly optimized,that is,the trained model is pruned with the channel pruning algorithm,the structure and parameters of the model are reduced under the premise of ensuring the accuracy,and the scale of the simplified network model is effectively compressed,so that the YOLOv4 network model can quickly detect and identify pedestrians without losing a large amount of detection accuracy,and the new YOLOv4 network model occupies less storage space and faster detection speed.