Computer Vision in the Sky: Ultralytics YOLOv8 and Deep SORT Synergy for Accurate Vehicle Speed Monitoring in Drone Video

Abstract

In recent years, the utilization of UAV video for traffic monitoring has experienced a significant surge in both popularity and effectiveness. This upward trend can be primarily attributed to its distinctive advantages, including flexibility, traceability, easy operation, and the wealth of information it offers. The present study introduces a comprehensive methodology tailored for the detection and tracking of vehicles in aerial footage, with a focus on determining their speed. This approach harnesses the capabilities of the Ultralytics YOLOv8 model and the deep SORT algorithm, aiming to establish a robust correlation between the detected vehicles and the drone's height above ground level (AGL). To precisely calculate vehicle speed, the study integrates a combination of well-established techniques. This includes addressing radial distortion through a higher-order distortion coefficient (k3) in the lens distortion correction process. Additionally, the study employs an Image to Real-world coordinate mapping approach based on a hybrid method of Horn-Schunck and Lucas-Kanade. Finally, the speed of identified vehicles is calculated using the Centroid point based geo-referencing techniques.

To ensure the precision of the proposed approach, a field experiment was conducted, capturing 9000 frame images from a test vehicle equipped with high-precision GPS. The experiment involved twenty groups with varying heights (ranging from 70 m to 100 m) and operating speeds (ranging from 7 m/s to 20 m/s, equivalent to 25 km/h to 72 km/h) over a 5-minute period at 30 frames per second.

The results obtained underscore the robustness and reliability of the proposed approach, as evidenced by a 97.19% precision in tracking vehicles and a 93.59% accuracy in object detection. Furthermore, the absolute and relative errors of the extracted speed remain below 1.7%, showcasing the high accuracy of the approach in speed estimations. The overall precision of the extracted parameters achieves an impressive 98.6%. These findings emphasize the efficacy of the proposed system in advancing traffic monitoring capabilities through the utilization of UAV video technology.