| In recent years,with the continuous development of information-physical fusion system(Cyber-physical system,CPS)related technologies,the corresponding security issues of intelligence has also become a key research direction.Many scholars at home and abroad in the field of target tracking extensive research,but for the typical application of CPS monitoring security scenarios in intelligent transportation,there is a large algorithm model difficult to deploy in embedded devices,and in obscured and dense scenes,common tracking algorithms In this thesis,the target tracking algorithm is used to track the target.To address these problems,this thesis divides target tracking into two stages: detection and tracking,and improves the target detection and tracking algorithms in terms of compressing the target detection algorithm model and improving tracking accuracy,respectively,and proposes a detection-based multi-target tracking algorithm:In order to explore the lightweight bottleneck performance of the detection algorithm for the problem of complex target detection algorithm model difficult to deploy in the surveillance and security environment in the CPS intelligent transportation system,this thesis proposes a target detection lightweight index LS＿Yolov5 based on CPS surveillance and security to evaluate the lightweight performance and degree of the algorithm.The YOLOv5 model is compressed by using a compact network structure that minimizes the change in the performance of the original model.By studying the efficient compact network characteristics such as Mobilenet network,Ghost Net network and Efficientnet network and the functional characteristics and structural details of each submodule in the YOLOv5 model.The YOLOv5-EBGN algorithm based on fused compact networks is proposed by fusing the Efficientnet network,which is efficient in backbone network feature extraction,and the Ghost Net network,which performs well in the neck network,and the network model size is reduced to 66% of the original one with only 1%reduction in accuracy and excellent performance in LS＿Yolov5 index.And the model size is only 9.4M more suitable for deployment in resource-constrained CPS monitoring security under the intelligent transportation system scenario.For the CPS intelligent traffic system complex scenes target tracking algorithm accuracy is not enough and in the dense occlusion scene tracking target omission problem.Based on the CIOU metric approach,for the problem of gradient disappearance due to insufficient smoothness of the aspect ratio indicator function,this thesis selects a smoother Soft Sign activation function with faster computation speed and proposes Soft＿CIOU,which improves the matching ability of Deep SORT tracking algorithm.And the original NMS algorithm for screening candidate frames is replaced by Soft-NMS algorithm,which makes the score change smooth and has stronger robustness by retaining more candidate frames.The model’s ability to perform on dense targets is increased.The proposed YOLOv5-EBGN+Deep SORT-SE target tracking algorithm improves the tracking accuracy by 10% compared with the original tracking algorithm,and achieves accurate tracking results in complex scenes. |
