The exceptional maneuverability of unmanned aerial vehicles (UAVs) has resulted in their utilization in diverse computer vision applications, particularly in intelligent traffic monitoring, where it improves efficiency and streamlines the procedure. However, there are still several obstacles that need to be addressed to automate these processes. One important task is the extraction of the foreground (vehicle) from complex traffic scenes. This research introduces a new vehicle recognition and tracking system for autonomous vehicle surveillance. The method utilizes U-NET to segment the aerial images. Following the segmentation process, we used the YOLOv8 deep learning system, known for its effectiveness in recognizing small-sized objects, specifically for vehicle detection. Moreover, to track several vehicles across frames of images, a method using SURF characteristics has been used for identification, assignment, and recovery. Ultimately, the tracking of these vehicles was accomplished by using DeepSORT, a technique that integrates the Kalman filter with deep learning to provide precise outcomes. Our suggested model has shown a remarkable rate of vehicle recognition and tracking during experimental validation. Our suggested model has acquired an accuracy score of 0.88 and 0.85 across the UAVID and AU-AIR datasets for vehicle detection. However, for the tracking of vehicles, the accuracy is 0.89 across the UAVID dataset and 0.84 for the AU-AIR dataset.
Vehicle Surveillance Using U-NET Segmentation and DeepSORT Over Aerial Images
23.05.2024
1483404 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch