Research And Achievement Of People Counting Algorithm In Video Based On Deep Learning

In the prevalence of big data,all walks of life are required to have the most comprehensive and accurate data,so that their own decision-making more forward-looking and scientific.Counting people in the surveillance video can provide a rational basis for the construction of public facilities so as to avoid wasting materials and reduce or avoid the safety problems caused by dense crowd in some public places.How to effectively count and manage the number of people in public places and the distribution of crowd is an urgent problem to be solved.One of the effective methods to solve this kind of problem is people counting technique with image detection.It used the extracted feature to judge where pedestrians are,and then count the number of pedestrians.The traditional people counting algorithms is difficult to achieve high accuracy in complex scenes.Although the accuracy of the people counting methods based on deep learning has been greatly improved,but it basically can't achieve real-time,and its accuracy rate is difficult to meet the practical application requirements due to perspective distortion and severe occlusion between pedestrians.In this paper,a people counting algorithm based on the deep learning detection model and tracking matching is proposed.In the case of multiple complex surveillance scenes,it can also ensure that the results of people counting are accurate and real-time.The main work and contribution of the thesis are:(1)In this thesis,we have a detailed analysis of current popular people counting algorithms and find out the problem in it.We use a method to counting people based on deep learning head-shoulder detection and tracking fusion.Firstly,we train a Faster R-CNN head-shoulder detector with Zeiler model to detect people with multiple poses and views.Next,we employ kernelized correlation filter(KCF)to track the people.Finally,we use fusion strategy to analyze the results of detection and tracking,and obtain a continuous and stable trajectory for people counting.(2)Based on the original Faster R-CNN object detection framework,three improved schemes are proposed to improve the detection accuracy for the head-shoulder detection task in surveillance scenes: The recall rate of the proposal in RPN network is improved by using positive and negative sample proportional equalization.Modifying the original frame structure,and we use the global context information(rather than local information)to determine the category of proposals,so it can reduce the false detection.Training model with online hard example mining methods to accelerate the model convergence process and reduce the error rate of algorithm.(3)The high-reliability,high-precision and high-speed kernel correlation filtering algorithm is used to track the detected head-shoulders.And the scale-adaptive kernel correlation filter tracker is used to solve the problem that the scale of objects is changed caused by perspective distortion.According to the characteristics of crowd in surveillance scene,a matching fusion method based on Hungary is used to analyze the results of head-shoulder detection and tracking.(4)In order to realize counting people with real-time by our people counting algorithm,a scheme of dynamic frame detection and matching is proposed.Which heavily reduces the running time of the algorithm with sacrificing a little accuracy.In this paper,the experiment and comparison of the head-shoulder detection algorithm and the overall people counting algorithm show that the proposed method can complete the task of people counting in the complex surveillance scene in real time and accurately.