Research on Real-time Detection and Tracking Algorithm for Low Slow Small Targets Based on the DeepSort

Aiming at the problems of large tracking errors and high loss detection rates in existing algorithms for detecting and tracking low slow small targets such as drones, an improved online low slow small target detection and tracking algorithm based on YOLOv5 and DeepSort is proposed. Firstly, the collected degraded image is processed to enhance the feature information of small targets; the YOLOv5 algorithm is used to detect the target, and the target motion model is established after initializing the Kalman filter. The target appearance model is established based on the appearance characteristics of the target detection results; the DeepSort algorithm associates the new detection result with the existing trajectory based on the data obtained from the appearance model and the motion model and performs the association measurement through the Hungarian algorithm to complete the matching of the trajectory and detection. At the same time, when detecting and tracking in the next frame, the ROI is set to improve the detection efficiency and tracking stability of small targets to achieve real-time detection and tracking of the low slow small target. The experimental result on the self-built dataset shows that the photoelectric equipment using this method can detect and track low slow small targets effectively.