Research On Classification Decision-making Technology Based On Target Recognition

The target recognition system equipped on the airborne platform has the advantages of wide application range,strong maneuverability and high monitoring efficiency,so it was a dual-use technology for both military and civilian using.Supported by the Aviation Science Foundation Project,the paper carries out research on the application technology of miniaturized and lightweight target recognition and task decision-making based on embedded platform to deal with the problems such as the limited information processing hardware resources of the airborne photoelectric observation system and targets recognition difficulty in the geodetic background environment.Meanwhile,the technologies of deep learning network optimization,deep learning target recognition,embedded platform implementation technology and task decision-making were studied.The main research contents were as follows:（1）The comparative analysis of a single-stage target recognition deep learning network framework and some constituent units were performed.The main network modules affecting the target recognition results were analyzed,and the feasibility and technical approaches that change the network structure were discussed to find a way to improve target recognition performance.And then the target recognition performance of several typical networks on public data sets was analyzed to select a high-precision scheme for air-to-ground target observation based on deep learning network.（2）A new data sets of air-to-Earth observation target recognition was built.According to the idea of FPN,a three-layer target feature vector scale based on the lightweight network model and the RFB module was put forward to deal with problems of feature extraction network shallow of the YOLOv3-tiny network,insufficient extracted semantic information,few target recognition feature vector scale and incomplete information,and to enhance the ability to extract target features under the ground background.Algorithm transplantation and performance testing of the optimized deep learning network on the TX2 embedded platform were done at the same time.（3）According to the decision-making mission requirements of the collision possibility of flying targets near the ground,the monocular vision observation system model,the position parameters between air to ground long-distance observation targets were applied to solve the relative position relationship between targets and between targets and cameras.And a fuzzy reasoning system and a fuzzy reasoning rules were designed to evaluate the collision risk assessment level of the target.Meanwhile,the ID3 algorithm was used to build a decision-making model.Lastly,the test experiment and data analysis of the decision-making model were finished.