Design And Implementation Of Outdoor Target Tracking Flying Robot

Target tracking is one of the important research directions in the field of computer vision.The technology of UAV has developed rapidly and widely used in various industries.The combination of UAV and computer vision algorithms can realize the active recognition and continuous tracking of ground targets,which is of great significance in military reconnaissance,public security and other fields.Aiming at the specified pedestrian recognition and tracking task based on the airborne sensor,the thesis mainly constructs the flying robot system and designs the target recognition and tracking algorithm.The verification experiments are carried out on the public data set and the self-built aerial photography data set.The main works of the thesis are as follows:Firstly,a flying robot system is designed.The system uses the four-rotor structure as the flying robot body,and applies the Pixhawk flight control as the flight control unit.The Jetson TX2 embedded processor and a HD camera mounted on the flying robot are used as the image processing module of the system.In addition,the software architecture and workflow of the outdoor target tracking flying robot are designed.Secondly,a method of pedestrian identification based on aerial image is designed and implemented.Tiny-YOLOv3 is used to detect the pedestrian in aerial image and the corresponding face.By cropping and aligning the detected face image,FaceNet is used for face verification,which can be used to determine the target pedestrian.The aerial image has certain particularity which is shot from the look-down angle.In order to improve the robustness of the training model,the aerial data set contains different flight altitudes and relative distances.And the two categories of pedestrian and human face are labeled.The Tiny-YOLOv3 target detection model is trained.Face images are cropped and then trained for classification.The self-built test set is used to verify the pedestrian recognition method.Thirdly,aiming at the tracking failure problem caused by target occlusion,a fDSST(fast Discriminative Scale Space Tracking)based on SSDA(Sequential Similarity Detection Algorithm)is designed to improve the ability of anti-occlusion for the target tracking.The oscillation degree of the response graph of the correlation filter is used to determine whether the target has be occluded.The model updating strategy with high confidence is adopted,and the SSDA algorithm is used to re-detect the missing target.The verification experiments are carried out on the public data set and the self-built aerial photography data set.The experimental results shows that the improved algorithm effectively improved the occlusion resistance ability compared with SAMF(Scale Adaptive with Multiple Features tracker),SRDCF(Spatially regularization Correlation Filter),and fDSST.Finally,the flight robot system is designed,and the target recognition and tracking algorithm is deployed to the flight robot platform.The actual flight test shows that the flying robot can realize the recognition and tracking of the target pedestrian.When the target is occluded by other pedestrians and recovered from the occlusion,the flying robot can still continue to realize the tracking.