Video People Counting Algorithm Based On Object Detection

With the widely used of video surveillance system,visual information has become a key research element in the security technology of modern security.Researchers have extended the technology in the field of computer vision to intelligent video surveillance systems,making the computer analysis the video information.People can obtain regional people numbers and population distribution through video surveillance,which can help enterprises to make operational decisions or help teachers to perform classroom attendance.At present,the real-time population statistics of video surveillance mainly rely on human eye recognition,the labor cost is high,the recognition efficiency is low,and video information cannot be understood and processed timely and effectively.The use of deep learning algorithms for demographics can help people solve complex repetitive tasks,which not only improves work efficiency,but also reduces labor costs.In the real scene,human body detection faces diverse environments,which exists some influencing factors such as illumination changes,object occlusion,and large object size differences,and the video is composed of thousands of image frames.The composition also has high requirements for the detection speed of the algorithm.This paper mainly based on the deep learning object detection method for human body detection and statistics of the final number.By analyzing the advantages and disadvantages of the existing feature extraction network,this paper proposes a parallel feature extraction sub-network based on dilated convolution,which makes the algorithm more robust for video people counting.Based on the difficulties of video people detection and the characteristics of video itself,this paper proposes a video-based people counting method.By researching on the existing methods of people counting,this paper designs video-based people counting algorithm.In view of the scale diversity problem in the people counting,this paper proposes three parallel sub-networks based on dilated convolution,which makes the extracted features retain more semantic information and alleviate the false detection caused by scale inhomogeneity in object detection.Aiming at the problem of object occlusion in the people counting,this paper designs a more distinguishing loss function,which solves the problem of missed detection caused by crowded occlusion and improves the recall rate of dense people detection.In view of the characteristics of video object detection,this paper adopts sequence-based non-maximum value suppression algorithm in post-processing,and uses the object of high-scoring in adjacent frames to improve the weak detection problem of the same object in the same sequence.This paper evaluates our method by experiments on several challenging people detection datasets,such as PASCAL VOC,COCO and Crowd Human data sets.This paper also performs experiments on ICG data sets for post-processing improvement of video.The final average precision is 88.4% on PASCAL VOC data set,and the speed is fast.By comparing with existing methods,the results sufficiently demonstrate the effectiveness and efficiency of the proposed method.