Lightweight And Embedded Implementation Of Target Detection Algorithm In Land Battlefield

As the earliest and most frequent battlefield space for military activities,the land battlefield is one of the important combat scenarios,which is crucial to the success or failure of military strategy and ensuring national security.With the improvement and development of science and technology,military target detection technology is also gradually optimized,and for the traditional detection methods in the land battlefield environment,there are time-consuming and labor-intensive problems,so military intelligent detection algorithms based on artificial intelligence,machine vision and other fields came into being.At present,the training and detection of most existing models rely on mainframe computers or servers,and the relevant algorithm parameters are redundant,which is difficult to deploy on resource-constrained equipment,but military battlefields have the characteristics of serious resource constraints,and can only be deployed on devices with small computing power,power consumption,and small size.Therefore,taking Person,Tank and UAV as detection targets,this paper carries out research on lightweight and embedded implementation of land battlefield target detection algorithms,and completes the lightweight and detection model deployment of target detection algorithms in military scenarios on the basis of analyzing domestic and foreign target detection technologies,which greatly reduces the cost of military target detection and provides decision-making basis for the rapidly changing battlefield situation,and the main research contents are as follows.(1)In light of the serious lack of corresponding data sets in the special scene of the land battlefield,this article completes the building and expansion of the data set,and evaluates the quality of the constructed data set.First,the COCO dataset was used to pick images with pertinent features,followed by web crawlers,movie and television frames,and finally the basic dataset.Second,the conventional data augmentation algorithm is applied in light of the issues with the single Angle scene,uneven distribution,and bad image quality in the basic data collection.There will be missed detection in the detection process due to the tiny size and obscure characteristics of the UAV in the object to be detected.We investigated the Cut Mix amplification method.The Mosaic data enhancement optimization algorithm is investigated in order to address the issues that the land battlefield has powerful invisibility,the detection target difference is relatively small,and the target between the two frames of the basic data set images is comparable.The expansion of the data set was finished using the three techniques mentioned above.Finally,the data sets before and after amplification are rationalized and verified using YOLOv5-s.The outcomes of the experiment demonstrate that amplification raises the caliber of the data collection.(2)This study carried out target detection network optimization and lightweight design due to the issues of target occlusion and camouflage,variable target scale,and large algorithm model size in military scenarios.First,to increase the model’s detection accuracy in complicated scenes,a channel and spatial attention method is added to the backbone network.Second,a modified Bi-FPN bidirectional feature pyramid is suggested to increase the receptive field,thereby significantly enhancing the ability to identify multiple scales of targets.Then,in order to decrease the number of model parameters,a better backbone network design built on Mobile Netv2 is suggested.Finally,the related channel pruning work of the entire network is finished in order to further reduce the model’s volume based on the improved network model discussed above.(3)This paper examines the embedded devices available on the market in light of the issue of low detection speed and accuracy in the current deployment of embedded platforms,and ultimately chooses TX2 for algorithm deployment and accelerated design.The target detection algorithm is initially deployed using the Py Torch framework after the necessary environment has been created on the embedded device.Second,the Tensor RT-based acceleration trial for the detection model is finished.Finally,this article completes the PYQT-based design of a related upper computer,improving the usability of human-computer interaction.The constructed MF-YOLO-0.8 field target detection algorithm,compared with the same type of YOLOv5 algorithm,improves the detection speed of the model and reduces the volume of the model while slightly reducing the accuracy,which can quickly and accurately detect military targets and analyze the situation with data.Relevant military background videos were used to verify the effect.