Real-time vehicle speed estimation using Unmanned Aerial Vehicles for traffic surveillance

Abstract

Drones are an emerging tool for traffic surveillance; however, they inherently lack the capability to solely obtain vehicle speed on the road. This Bachelor's thesis presents the design, implementation and study of a system to detect the position, velocity and type of vehicles using the video stream obtained from drones. The solution is created to be used with any kind of aerial vehicle but is tailored for the drones in the European project LABYRINTH, of which the thesis has been a part. The tool utilizes the video feed from a sole camera and the telemetry data from the drone to detect, track and project the objects present on the road from the image into reality. This allows for an estimation of their position and speed. The detection and tracking algorithm implemented is the Simple Online Real Time algorithm, which is often referred to as SORT. Once the position has been acquired, another stream is generated that displays the same video, but with the bounding boxes, velocity and confidence ratings of all identified vehicles, with an overall computing time lower than the frame rate. After implementation, the tool underwent testing in a simulated environment to determine its assets and shortcomings, and was used during the LABYRINTH traffic monitoring flight tests. The Bachelor's thesis achieves the aimed objectives with minimum resource utilization, using readily available logic and open-source software to strike an optimal balance between real-time functionality and precise detection of vehicle position.