Research On Battlefield Target Detection And Tracking Based On Deep Learning

In the future,battlefield operations will develop towards the direction of information and intelligence.It will speed up battlefield operations and process a large amount of information in a short time.Thus,it is necessary to realize real-time monitoring of battlefield targets.The complex combat environment and the characteristics of tank target movement make the research of battlefield tank target detection and tracking system a challenge.Therefore,the development of automatic detection and tracking technology of tank target in different scenarios has important application value.In the process of tank target detection on the battlefield,the factors affecting tank target are analyzed,such as similarity to the background,the presence of occlusion,different scales,different faces,dust and smoke.And two kinds of problems exist in the process of tank target detection are obtained:the detection problem when tank target has occlusion and the detection of small tank target.In this thesis,Faster R-CNN algorithm and migration learning algorithm are used for target detection.Besides,in order to prevent the tank target prediction region from appearing on the non-target area,the Repulsion Loss(RepLoss)function is introduced to improve the detection accuracy when there is occlusion.To solve the problem of small tank target detection,the method of optimizing candidate region is adopted.The loss function of candidate region on conv3-3 and conv5-3 layers is compared to optimize candidate region,at the same time,the sliding window of the Region Proposal Network(RPN)adds a 642-size anchor to enrich the number of candidate regions and optimize the quality of the candidate regions.Validation experiments are conducted on the tank target test data set to compare the detection performance of this algorithm and other algorithms in different scenarios.The results show that the Faster R-CNN algorithm based on the optimized candidate regions has a high detection accuracy under the influence of small tank target and occlusion.Based on the deep learning results of tank target detection,the tank target tracking system is constructed.In this thesis,there are three kinds of problems in the process of tank target tracking:similar interference,occlusion recovery when tank target is occluded and restoration when tank object tracking is wrong.In the framework of particle filter tracking algorithm,the Kalman filter is used to update the state of particles,and the importance sampling of particles is guided to reduce the error between particle state and target real state.The overlap rate coefficient is used to judge the occlusion and Kalman prediction is used as the target state to update the template to solve the problem of occlusion recovery when there is occlusion in the tank target tracking process.The improved resampling algorithm is used to solve the small weight particle degradation problem,which increases the number of small weight particles and the diversity of particles.In order to verify the effectiveness of the improved particle filter tracking algorithm,the performance of different algorithms is carried out for three kinds of problems in tank target tracking.The results show that the improved particle filter tracking algorithm can effectively solve the problems in the process of tank target tracking,and also achieve a good tracking effect.