Multiple Pedestrian Tracking Based On Deep Learning

Multiple pedestrian tracking technology has attracted more and more attention in recent years because of its huge commercial value.The traditional multiple pedestrian tracking method is only suitable for scenes with a single environmental background and a small number of pedestrians.In practical application scenarios,the environment backgrounds of pedestrians are usually very complicated,pedestrians are crowded and blocked,and pedestrian postures are constantly changing.It is challenging to research multiple pedestrian tracking technology.Recently,deep learning technology has shown excellent performance in various research directions of computer vision.This paper studies the multiple pedestrian tracking method based on deep learning technology,using deep learning technology to detect and track pedestrians in video or image sequences.Firstly,an pedestrian detection model based on deep learning is used to detect pedestrians,and then a convolutional neural network is used to extract the apparent features of pedestrians,combined with pedestrian motion feature to associate pedestrian identities in different images to achieve the purpose of multiple pedestrian tracking.The multiple pedestrian tracking method studied in this paper mainly includes three parts,and the specific research content is as follows:1.Research on pedestrian detection methods using deep learning.For multiple pedestrian tracking,the first step is to obtain the position of the pedestrian.This paper studies a pedestrian detection model that can accurately output the position of pedestrians in a scene with complex environmental background,high density and serious occlusion.Because the two-stage object detection model has higher accuracy,the Faster R-CNN is used as the framework in this paper.In the pedestrian detection model,using ResNeSt101 as the backbone network to enhance the model's feature extraction capabilities in complex environments,while adding the FPN module to enhance the model's ability to detect smaller pedestrian objects,and adding BN layer after each convolutional layer of FPN accelerate network convergence and improve the model's robustness.Using deformable convolutional networks in the backbone network to enhance the ability to extract irregular features of pedestrians.In view of the dense occlusion of pedestrians,an iterative detection method is introduced into the detection model to output more pedestrian positions through multiple iterations of detection.The balance between accuracy and speed can be achieved by setting the number of iterations for different degrees of pedestrian density.The regression loss function of the model is improved,and the repulsion loss function is used to solve the dense shielding problem in pedestrian detection.In this paper,we analyzed the deficiencies of the repulsion loss function and made improvements.The bound repulsion loss function was called which improves the ability to detect crowded pedestrians.2.Research on the pedestrian appearance feature extraction model using deep learning.In multiple pedestrian tracking,the appearance features of pedestrians are used to measure the similarity between different pedestrians.Aiming at the problems of poor pedestrian image quality,pedestrian occlusion,and similar characteristics of pedestrians,a multiple granularity network is used to extract the global features and local features of pedestrians.At the same time,for the number of parameters and speed problems,a lightweight convolutional neural network is used as the backbone network of the model.In this paper,the fine-grained quantity,the retained pedestrian feature dimensions and the fine-grained division method are experimentally studied.The network designed in this paper named M-MGN which can maintain high performance with greatly reduced parameters.Using M-MGN model to extract the appearance features of pedestrians can effectively reduce the number of pedestrian ID exchanges.3.Research on the method of multiple pedestrian tracking.After detecting the position of the pedestrian obtained by the model,M-MGN was used to extract the pedestrian appearance features.Aiming at complex crowded scenes,feature smoothing operation is adopted for pedestrian appearance features to increase the matching accuracy in the case of occlusion.Then,the similarity between the pedestrian and the track was measured by combining the motion features,and a data association method was designed to match the pedestrian detection results with the track.Then new pedestrian tracks are obtained and updated.This paper verifies the performance of the detection model on the CrowdHuman dataset,verifies the effect of pedestrian feature extraction on the Market1501 dataset,and verifies the multiple pedestrian tracking effect of the model studied in this paper on the MOT dataset.On the MOT16 dataset,MOTA scored69.3 points,and on the extremely dense pedestrian MOT20 dataset,MOTA scored64.4 points,indicating that the model studied in this paper is not only suitable for dense scenes,but also suitable for simple tracking scenes.