Single-person Detection And Tracking In Unmanned Aerial Vehicles

Online single-object tracking is an important problem in computer vision,which is widely used in intelligent video monitoring and human interaction.With the popularization of commercial drones,application scenarios,such as suspect tracking and user follow-up,are derived.Online single-person tracking(OSPT)in drones has become a primary requirement for unmanned-aerial-vehicle(UAV)intelligence.Given the person’s position in the first frame,intelligent following should track the object in subsequent frames,which will suffer from occlusion,long-period disappearing,etc.Meanwhile,a UAV should adjust its flying direction and angle,so it requires real-time performance of tracking algorithms.Unfortunately,there is as yet no online specific person tracking dataset from a low-altitude UAV.It is important to establish such a dataset and an algorithm for OSPT to realize robust tracking.First,we establish a dataset towards an OSPT subject,named UAVP100,which contains 100 tracking videos,nearly 13 w images,and 11 different challenges.We compare the aspect ratio,object relative size,and challenges of UAVP100 with other datasets to analyze the attributes of UAVP100.Meanwhile,we also evaluate 20 state-of-the-art tracking algorithms in UAVP100,which also shows that the proposed UAVP100 is more challenging.In addition,we research relative object detection algorithms based on deep neural networks and compare their performance in UAVs to find the difference between UAV scenes and general scenes.In addition,we analyze the tradeoff between different detection algorithms after fine-tuning by a UAV person-detection dataset to obtain the most suitable one for UAV application.Second,we propose a tracking framework for OSPT that contains three modules,tracker,verifier,and detection-reid to deal with long-term disappearance and re-appearance and other problems.We track the position of the object by a correlation-filter-based tracking algorithm,and adaptively update the template.The verifier adjusts the confidence of the tracker by the response map,and the detection-reid module performs person detection and re-identification to re-locate tracking objects.We compare our framework with the other 20 algorithms by multiple qualitative and quantitative experiments,the results of which prove that our algorithm can effectively balance tracking robustness and time complexity,and realize long-term and large-scale real-time robust tracking of a single person with an UAV.