Research On Pedestrian Stay Detection Algorithm In Case Handling Area Based On Deep Learning

With the rise of intelligent monitoring,the research on video analysis is increasing in all walks of life,and pedestrian stay detection algorithm is one of the research directions.The purpose of pedestrian stay detection is to successfully detect pedestrian stay behavior by according to the monitoring video to get pedestrian movement tracks,and then calculating activity time in a certain area.Because of the rapid development of computer computing power,which makes the advantages of strong feature expression ability and simplified mathematical operation are rapidly emerging,deep learning has gradually become one of the mainstream methods to deal with problems in the field of computer vision.In this thesis,we design a more accurate and adaptable pedestrian stay detection algorithm PSD,which is composed of RD-DeepSort tracking algorithm,CLABD-Net re-identification algorithm and stay discrimination algorithm by introducing the deep learning method.The specific research contents are as follows:(1)Improved tracking algorithm RD-DeepSort.DeepSort consists of detector,feature similarity calculation and cascade matching.Firstly,the detector is selected by test three detection networks in detection performance and DeepSort,which has three detection networks respectively,in tracking performance.The Refine Det net with balanced performance is selected as the detector of DeepSort.Secondly,the similarity measure of motion model is the Mahalanobis distance between the motion position estimated by Kalman filter and the detection box generated by the detector in DeepSort.On the one hand,because the two parts of Mahalanobis distance come from two systems respectively,that is,they obey different distributions,it cannot be considered that the measure obeys chi-square distribution.On the other hand,only the covariance matrix of Kalman filter is considered when calculating the covariance matrix of Mahalanobis distance witch ignores detection algorithm.Based on the above two points,the calculated value of the similarity measure of DeepSort motion model is larger than the actual value.So this thesis proposes RD-DeepSort which reduces the measure to reduce the number of false negative examples.In this algorithm,the difference between the track and detection box of the previous frame is used to make up for the influence of the detection error of the current frame on the motion model.RD-DeepSort test result shows that MOTA and MOTP are improved by1.11% and 2.43% respectively.(2)Designed pedestrian stay detection algorithm PSD.Because the track ID cannot correspond to the actual pedestrian and ID switch phenomenon occurs frequently in tracking algorithm,it is unable to calculate the pedestrian stay time in a specific area.So track ID recognition is a difficulty in pedestrian stay algorithm.In this thesis,the person re-identification is added to the stay detection algorithm to make the track ID correspond to the actual pedestrian to solve the difficulty.Firstly,putting forward CLABD-Net by changing the loss function in re-identification algorithm to improve the accuracy of re-identification.Secondly a pedestrian stay behavior detection scheme is designed according to the actual using scenario and the available data.Based on this scheme,experiments and improvements are continuously carried out on the self-built data set in the scene of case handling area.Finally,we can get PSD with excellent accuracy and robustness by two optimizations of reducing decision threshold and frame extraction tracking.The test results show PSD not only has high accuracy,but also has good reliability.In the test environment,the test results show that accuracy and precision reach95.73% and 97.26% respectively,and the missed detection rate as low as 6.58% and false detection rate is 2.27%,which meet the demand of stay behavior detection in the indoor case handling area.