Research On Target Pedestrian Search And Detection In Edge Computing

Recent developments in the field of end-to-end detecting and reidentification(re-ID)have led to a renewed interest in person search,Similar to image search and face recognition technology,person search model takes the whole body image of target to search for which cameras the target appears.The most difficult part of person search is how to search from different cameras angles.Person search models are widely applied in shopping mall and suspect search.The state-of-the-art person search models apply deep learning as their backbone technique so that they consume huge GPU computing resources in particular,which scale would limited with the increase of cameras.There are more and more embedded devices having basic computing resources such as NVIDIA Jetson TX2 and Raspberry Pi.The cloud server's computing loading would be offload if we can use their computing resources.Moreover,the person Search models are one kind of multi-task model with the possibility of cutting and compressing them.This study combine a pedestrian detection model with a reidentification model and proposes a new architecture based on a one-stage pedestrian detection model.Although there have been related studies on the end-to-end person search models,they are all based on two-stage pedestrian detection networks,which resource consumption is large and difficult to achieve the purpose of real-time search even if the accuracy has been improved.The person search model based on the one-stage detection network can save nearly 2 times the number of parameters and increase the recognition speed nearly 3 times.The one-stage pedestrian detection network is affected by the imbalance classification problem causes lower accuracy than the two-stage detection model.Therefore,this study proposes a new training model method for this problem.Since the similarity calculated by the pedestrian re-identification subnet is not affected by the imbalance classification problem,it is the good reference to annotate how important of each training sample.Our training method employs an auxiliary parameter based on the similarity to the pedestrian detection subnet to control the importance of each training sample.With the importance of different simples,a dynamic loss function is provided,and the accuracy of the pedestrian detection subnet in able to match and surpass the state-of-the-art ones.Edge computing is a research on how to utilize the computing resources of edge devices to collaborate with the central server.This study refers to the research results so far of the most advanced edge computing technique to improve computing-intensive tasks such as person search to scale up model inference,and proposes a model segmentation and computing task allocation.In the part of model segmentation,the end-toend person search model in this study is compressed so that multiple segmented models can be deployed on edge devices with limited hardware resources.In addition,the strategy of uploading and offloading is based on the principle of maximum entropy.After adding multiple restricted conditions,the model with the maximum resource utilization rate is selected with the minimum probability of crash of the edge device.We apply CUHK-SYSU and PRW pedestrian search training set for training,and set up a scene similar to this data set for experiments and analysis.Model recognition speed,accuracy,and training convergence speed are all the focus of experimental analysis.The experimental results show that the edge computing and the lightweight person search model save the server 80% of computing resources and improve the recognition speed by 30%.In addition,the recognition accuracy is matched with the state-of-the-art person search model NPSM.