| In recent years,the arrival of the 5G era has brought unprecedented opportunities and challenges to the artificial intelligence industry.As the most important field in artificial intelligence,computer vision has played an indispensable position and role in intelligent security.Person recognition and object tracking are two important components of computer vision and the most important part of intelligent security systems.In recent years,they have been extensively studied and achieved considerable research results.Unfortunately,as the research on person Re-ID and object tracking continues to deepen,some difficult issues also slow down the research on these two tasks.In terms of person re-identification,although most studies have achieved a high recognition effect,in the case of low light,the recognition effect will be greatly disturbed due to the loss of pedrson color information,which limits these research results application in actual system.In terms of target tracking,most of the current tracking methods cannot train a good target recognition module,and when the target movement speed is too fast,it may easily lead to the situation that the target jumps out of the bounding box generation area which may cause tracking failure.In order to deal with these difficult problems in two research directions,this paper proposes a cross-modal multi-granularity feature fusion network,and a single target tracking method of fine-grained networks based on dynamic programming for target tracking.The main contributions and work content of this paper are summarized as follows:1.For person re-identification,Considering that more and more cameras have infrared camera functions,this paper has conducted person Re-identification research under the cross-modal conditions in the field of RGB-IR.This paper proposes a cross-modal multi-granularity feature fusion network.This network uses a newly designed dual-stream attention unit constraint network to prompt it to learn discriminative shared fine-grained features on person between different modalities,and combine it with the global features,and finally achieve an excellent recognition effect.2.For Single Object Tracking,this paper introduces Dynamic Programming based on Fine-grained Network(DPF)to solve the above shortcoming.In anutshell,this paper has made the following contributions: 1)The Io U value is estimated through the Io U prediction network,and different targets in various search areas are dynamically selected through the estimated IOU value.2)An Res-attention mechanism is added to the Resnet structure to ensure that the network constantly pays attention to the areas of interest in the picture and improves the discriminability of feature extraction.3)A new online tracking method based on dynamic programming is proposed to solve the problem that the target jumps out of the bounding box generation area. |
