Research On Ground Target Detecting And Situation Assessment Of UAV

Research on unmanned aerial vehicle(UAV) in the field of artificial intelligence and applications receive increasing widespread attention. Among them, UAV detects, tracks ground targets and its global situation assessment are main points. UAV indoor location, rapid detection and tracking of ground targets are studied, and based on this, UAV's global situation assessment method for ground targets is researched.Firstly, the indoor UAV localization problem in time and accurately is researched. Light flow information is combined with visual landmark under Kalman filtering to Self-Localizate. RANSAC method is used to extract ground features, feature points to inverse projection transformation; match for the visual signpost, artificial landmark method is adopted to improve the positioning. For UAV slow monocular vision positioning problem, optical flow sensor information fusion Kalman filtering, in order to improve positioning speed.Secondly, the moving target detection method combined with support vector machine(SVM) is researched. Due to the flight velocity and attitude of UAV change frequently and quickly, its ground target detection requires high real-time performance, strong adaptability. Use the background modeling method to extract prospect targets as interest areas. Based on gradient direction histogram(HOG) characteristics, support vector machine(SVM) of multiple targets detector is introduced, to ensure the accuracy of the detection at the same time improve the real-time performance of targets detection.Thirdly, a multiple similar targets tracking algorithm based on improved hough forest framework is proposed. For the failure of monocular visual tracking multi-similar targets algorithm with influence factors such as occlusion, an improved online Hough forest tracking framework is presented to formulate tracking problem as a Maximum a posteriori problem to realize the continuous tracking of multi-similar targets. Through online Multi-objective samples collection and appearance and motion information extraction, Hough forest is constructed to associate multi-target trajectory by training for track association probability. Low-rank Hankel algorithm is employed to correct Trajectory, modify associated errors and improved the efficiency of online update of the training set. Tracking accuracy and robustness of the single camera vision are effectively improved within similar or inter-occlusion targets.Finally, global situation of objectives on ground interested area is estimated by the hidden markov model. Modeling analysis was carried out on the ground moving target, to extract the situational factors; The Baum- Welch algorithm is used to train parameters of hidden markov model which make prediction model is obtained; Application the model to forecast the multi-targets global trend, and provide decision-making basis for UAV searching and controlling targets.