’Low Small Slow’ Target Detection And Tracking Technology Based On Air Surveillance Video

‘Low-Small-Slow’ usually refers to low-altitude flight,small volume,slow speed unmanned target,which has been widely used in military and civil fields,such as anti-terrorist stability,geological survey,and security of important airspace and other scenarios.However,due to the characteristics of Air Surveillance Video and‘Low-Small-Slow’ target itself,there are many problems such as small sample size,small target size,and background interference under the Air Surveillance Video angle,which restrict the further improvement of detection and tracking performance to some extent,and the existing algorithms have poor adaptability for small target detection and tracking tasks under complex background.In view of the above situation,this paper presents a ‘Low-Small-Slow’ target detection and tracking technology based on Air Surveillance Video.Key technologies such as target background reconstruction data expansion based on diffusion model 、 lightweight single-stage small target detection based on multi-scale aggregation,and small target tracking based on enhanced network based on focus center feature are studied in turn.The specific contents of this paper are summarized as follows:Firstly,to solve the issue of insufficient samples in the target datasets of‘Low-Small-Slow’ from the perspective of air surveillance,this paper presents a target background reconstruction data augmentation method based on diffusion model.To solve the problem of missing label information from self-collected data,a semi-supervised model is trained to complete the remaining image labels by manually labeling some images as input.Separate the target from the background,and after data expansion through the diffusion model,the target and background are randomly reconstructed to restore scene complexity,providing a large amount of data support for subsequent detection networks.Secondly,in response to the problem of insignificant features of small aerial targets that make it difficult to detect,this paper proposes a lightweight one stage small target detection method based on multi-scale aggregation,which optimizes the network structure by clipping the output layer of the base network Yolo V5 that does not match the size of the small target in the air,avoiding the calculation of redundant parameters to increase the training speed of the model.A multiscale attention aggregation module is introduced to aggregate small and weak target information in the top-down path of the network to improve the accuracy of the detection network and effectively enhance the robustness of the subsequent tracking network.Thirdly,in response to the issue of target confusion and tracking loss caused by complex scenes from aerial perspective,this paper presents a small target tracking method based on the Focus Center Feature Enhanced Network.In the training phase,batch processing is used to ensure the consistency of input data to improve the real-time performance of the tracking network training,and a deeper residual structure is used to replace the feature extraction part of the base network Siamese FC to enhance the network’s perception of weak and small targets.The focus center information module is designed to mix channel information and spatial information by multiple convolution layers and three important branches without adding redundant parameters,so that the tracking network pays more attention to the target information,thus improving its anti-jamming ability,and ensuring the adaptability and stability of the tracking model.Finally,through the study of the above key technologies,this paper completes the training and tracking network construction of the ‘Low-Small-Slow’ target detection model based on the data set from the air surveillance perspective,and conducts qualitative and quantitative experimental analysis to validate the key technologies proposed in this paper.