Research On Pedestrian Detection And Tracking Algorithms For Complex Traffic Scenes

Autonomous driving technology is an important means to improve the efficiency of traffic operation,reduce the accident rate,and enhance the intelligent experience of assisted driving.With the continuous improvement of hardware computing power and the iterative update of low-cost environmental state acquisition devices,tasks related to autonomous driving have attracted extensive attention and in-depth research by researchers in recent years.An automatic driving system is usually composed of environment perception,path planning,and decision-making control.Achieving accurate perception of the state of complex traffic scenes is the basis and premise for the stable operation of system.Due to the advantages of machine vision-based environmental perception technology with a wide range of applications,convenient installation and deployment,high algorithm maturity,and relatively low hardware prices,it has gradually become one of the main research directions at present.This paper focuses on pedestrian detection,single-object tracking,and multi-object tracking in the environmental perception task of autonomous driving.By exploring and improving the existing problems and deficiencies in the relevant algorithms at current stage,the effective detection and trajectory estimation of pedestrians in complex traffic scene can be achieved.This can give early warning and avoid accidents.The main contents of this research are listed as follows:(1)To solve the problem that the part-based pedestrian detectors only use the spatial prior between the sub-model and the main model to encode the pedestrian model,and lack of the effective description of the associated attributes between the parts,an improved pedestrian detection framework based on a tree-structured graphical model is established.First,define the parent-child relationship between the various parts of the human body,and obtain the hidden variables of the training samples by clustering the parent-child parts to their types.Second,for the purpose of taking both intra-class tightness and inter-class differences into account,a hybrid particle swarm clustering algorithm containing a two-phase fitness function based on the mean square error and the David bouldin index is constructed.It is not only effective in estimating the number of optimal cluster centers,but also in eliminating the effect of random initialzation on the clustering accuracy.Finally,the offline training model is used in the detection stage,and combined with the dynamic programming algorithm to solve the state transition equation to generate the detection results.The experimental results on the benchmark dataset indicate the effectiveness of the proposed improved strategy,as well as the accuracy and robustness of the overall pedestrian detection framework.(2)To solve the performance limitations of the YOLO-V4 algorithm in dealing with dense pedestrian occlusion in complex traffic scenarios,an improved YOLO-V4 pedestrian detection framework based on optimized network structure,effective multi-scale feature fusion strategy and specific network loss function is proposed.First,design a single output network according to the pedestrian detection object,and use the proposed ladder fusion strategy to integrate image information from multiple scales,which is effectively to solve the problem of invalid anchor allocation to the approximate target,and ensures the anchor's aspect ratio estimation is still driven by data.Second,adjust the resolution ratio of the output feature map to the network input image to reduce the label rewriting cases of the training samples.Finally,the concept of repulsive force is introduced to optimize the bounding box regression loss function to further improve the robustness of the model to the detection of densely occluded pedestrians.The experimental results show that Im-YOLOV4 can significantly improve the pedestrian detection performance of the original algorithm,and it has higher robustness in the field of autonomous driving.(3)To improve the performance of the siamese network-based single object tracking algorithm in complex traffic scenarios,a multi-stage SOT framework D-CRPN(Siamese Tracking with Deeper Networks and Cascaded RPN,D-CRPN)is proposed.First,we exploit the cropping-inside residual units for reforming Res Net to break the spatial invariance restriction.Second,in order to integrate the deep semantics information and shallow spatial information effectively,specific Feature Transfer Block(FTB)is designed for different stages of the network.For the RPN outputs of the overall network,we use decided-level fusion strategy to further improve the tracking performance.Finally,a quality measurement method is proposed for the output response heatmap of the RPN module,and it is applied to the calculation of adaptive weights in decision-level fusion.The experimental results show that the proposed D-CRPN can utilize the complementarity of the output information at different stages of the network when using a deeper backbone network,and improve the object tracking accuracy effectively.(4)In view of the demand for multi-target tracking of people in actual traffic scenarios,and combined with the proposed Im-YOLOV4 pedestrian detector and D-CRPN single-object tracker,a multi-object tracking framework based on spatio-temporal cues fusion and optimized cascade matching is proposed.First,a reliable online multi-object tracking algorithm process is developed using the motion model and appearance model.Second,in the spatio-temporal cue fusion stage,the candidate results are expanded through the tracking quality evaluation,and the adaptive SOT stop update conditions are formulated considering the occlusion factor.In the long-term clue extraction stage,the trajectory historical appearance screening strategy and the trajectory scoring mechanism are proposed to improve the quality of long-term leads and optimize the priority order of cascade matching.Finally,motion estimation and motion compensation are used to eliminate the influence of camera shake on the effectiveness of motion information constraints,and the appearance and motion matching and data association algorithms are used to complete the multi-object tracking task.The experimental results on the MOT Challenge 16 and MOT Challenge 17 datasets show that the proposed method achieves robust multi-object tracking of pedestrians in real scenes by rationally fusing long and short-term video cues.