Battlefield Soldiers Based On YOLO And Related Filtering Detection And Tracking Research

In recent years,due to the rapid development of computer vision technology,traditional combat methods can no longer meet the requirements of modern battlefield operations,and thus a variety of popular artificial intelligence-based technologies have been applied to the military field,such as target detection and target tracking.The addition of these technologies has led to the emergence of intelligent weapons,thus enhancing combat efficiency.Over time,smart weapons have become more powerful in that they are able to effectively identify and strike specific targets.Therefore,the automatic identification and tracking of soldiers in the battlefield environment by various smart weapons like human brains has become an important challenge that needs to be addressed.In this paper,we focus on soldier targets in the battlefield environment,and use the improved YOLOv5 s detection algorithm and DSST tracking algorithm to conduct intelligent detection and tracking research on soldier targets in the battlefield environment based on the deep learning-based target detection method and the correlation filtering method with stable tracking efficiency.The following work was mainly carried out:1.The soldier target detection and tracking dataset in the battlefield environment is constructed for the problem of less soldier data in the battlefield environment: This thesis completed the soldier data collection through operations such as web crawlers and film screenshots,and the original data was amplified and defogged to improve the quality of the dataset for the problem of little collected data and large interference in the picture data,and the battlefield soldier image dataset was constructed.2.For the problems faced in the target detection process,such as complex battlefield environment,target scale changes,small-scale targets,and real-time,A battlefield soldier detection method based on improved YOLOv5 s is proposed: first,the backbone network is improved,and the improved lightweight network Ghost-Bottleneck module is introduced into the backbone network to reduce the network model computation and improve the model detection speed;then,the network performance is improved by introducing a convolutional attention mechanism into the backbone network.Secondly,the feature fusion network of the model is improved,mainly for the two problems of target scale variation and small-scale targets,and the multi-scale feature fusion network is reconstructed and the traditional convolution is replaced by self-correcting convolution to improve the detection accuracy of scale-varying targets and small-scale targets in the battlefield environment.Finally,the effectiveness of the improved model in the detection part of this thesis is demonstrated by ablation experiments and comparison experiments of the effect graphs before and after the improvement.3.A battlefield soldier tracking method based on the DSST algorithm is proposed to solve the target occlusion and multi-target matching loss problems during soldier tracking on self-constructed datasets.The method combines the TLD tracking algorithm,which can reduce the tracking drift phenomenon in the case of occlusion.Meanwhile,the Hungarian matching algorithm is used to correlate the data of different frames of targets to solve the multi-target matching problem and improve the success rate of tracking.The experimental results show that the tracking algorithm proposed in this paper performs better in dealing with the occlusion and multi-target matching problems during soldier target tracking compared with the benchmark method,and improves the tracking success rate and accuracy.