Pedestrian Visual Detection And Tracking For Rotor Flying Robot

The rotor flying robot are widely used in anti-terrorism,explosion-proof,traffic monitoring and earthquake relief because of their small size,strong maneuverability and easy control,and their ability to operate in extreme environments.The rotor flying robot need to track specific targets and transmit real-time information to ground stations.Therefore,the detection and tracking of rotor flying robot based on vision sensor has become a hot research direction.Firstly,a Kinect-based multi-aircraft visual tracking and positioning method for rotor flying robot is studied in this paper.Using Kinect-based external visual positioning system built independently by the laboratory,aiming at the occlusion problem of the rotor flying robot in multi-aircraft tracking and positioning,a segmentation multi-target tracking algorithm is designed,obtaining the image coordinates of the target.Then,according to the relationship between Kinect coordinate system and world coordinate system,the image coordinate is transformed into three-dimensional coordinate to realize the flight experiment of the rotor flying robot.The static positioning accuracy of the rotor flying robot is less than 5 mm,while the positioning accuracy is less than 45 mm under occlusion,and the frame rate is 30 fps.The experimental results show that this method can solve the occlusion problem very well.Based on this method,the application of pedestrian vision detection and tracking for rotor flying robot is studied.Secondly,this paper studies pedestrian detection method based on Faster R-CNN framework.For the occlusion problem of pedestrian detection,aggregate loss function is used to replace regression loss function in the first stage of this method,which makes the prediction frame closer to the real pedestrian area.In the second stage,a new occlusion pooling perception unit is designed,and pedestrians are divided into five parts for feature extraction and visibility score prediction.At last,the feature of the whole picture is weighted and fused,and then classified and regressed.The experimental results show that the proposed method is robust to pedestrian occlusion detection.Finally,this paper studies the pedestrian tracking method of rotor flying robot.Based on the airborne vision platform of rotor flying robot,aiming at the problem of blurring motion image caused by dithering of rotor flying robot and partial occlusion during pedestrian tracking,the pedestrian detected by pedestrian detection method is used as input,and the pedestrian nearest to image center is selected by Euclidean distance as tracking target;the tracking method combining HOG and color features is used to track pedestrian targets and obtain image coordinates.Then,the coordinates are converted to the angular and linear velocities required by the rotor flying robot by using the image-based visual servo algorithm to control the rotor flying robot to track the pedestrian target.The experimental results show that the method can complete the pedestrian tracking mission in real time and accurately.