Research on UAV Multi-Object Tracking Based on Deep Learning
Xi Luo, Rui Zhao, Xiang Yang Gao · 2021 IEEE International Conference on Networking, Sensing and Control (ICNSC) · 2021
UAV is widely used in civil or military fields due to its advantages of flexibility, compact and lightness, and it can replace human beings to explore unknown regions or perform various dangerous tasks, such as terrain survey, border patrol, intelligent transportation, power grid detection and disaster detection. Multi-object tracking is an important field in computer vision that has been studied for a long time. In recent years, with the rapid development of UAV technology, object tracking based on UAV has become a research hotspot. In this paper, the application of deep learning in UAV object tracking is studied based on the improved tracking-by-detection multi-object tracking neural network. The processed public data set is used to train the backbone network based on CSPDarknet53 as the detector while the dataset of cars is used to train a pretraining apparent feature vector based on deep learning. Besides, the Kalman filter is used to extract object motion information and update the prediction. Finally, tracking results are obtained by Hungarian matching algorithm. Several groups of experiments on the UAV123 data set show that the trained multi-object tracking algorithm on UAV platform can track the object stably under rapid object movement, fast turning of object motion, multi-object motion, object scale changes and other conditions.