Research On Person Locating And Counting System Based On Deep Learning In Surveillance Video

With the continuous development of society and technology,the demand for intelligent analysis system of surveillance video scene is also rising.Among all the applications of surveillance video analysis,the study of locating and counting people is important.Based on the theory of deep learning,this thesis designs a object detection model which can be used to locate and count people in the image by the method of multi-task learning.Combning with the post-processing algorithm,the accuracy of the system can be further improved.The main study of this thesis includes:1)This thesis explains the reason of adopting object detection model by showing the scene of deployed system,then collects and annotates the video data.Through semiautomatic annotation,the bounding box and mask annotation required for training can be quickly obtained.The bounding box can be used for the training of object detecion,and the mask can be for the training of instance segmentation and semantic segmentation.2)Based on the idea of multi task learning,instance segmentation and semantic segmentation are introduced to further improve the accuracy of human detection.In order to segment instance efficiently,this thesis introduces a additional mask feature branch and a controller branch to output the masks of different instances based on key point detection network.This thesis further introduces the semantic segmentation branch for learning.So far,this thesis designs a real-time multi task learning model.To locate and count the people in images,this thesis analyzes the best threshold of detection results through experiments,and evaluates the personnel counting and locating metrics under the best threshold,then quantitatively analyse the relationship between detecting performance and the locating-counting performance.3)This thesis deploys the heatmap propagation algorithm to further improve the locating and counting performance by making use of temporal information in video.The optimal hyper-parameters of the heatmap propagation algorithm are determined by sufficient experiments.The mask non-maximum suppression algorithm is introduced to drop the false detection.The experimental results show that the mask non-maximum suppres-sion algorithm has the advantages over the bounding box non-maximum suppression al-gorithm.4)Based on the multi task learning model and post-processing algorithm,this thesis designs an application system that can locate and count people in multi-channel video streams.The system can raise alarm to the personnel leaving the post,so as to better prevent the accidents caused by the unauthorized offpost.