Vehicle Speed Estimation Using YOLO, Kalman Filter, and Frame Sampling

Vehicle speed estimation based on video feed can be used to enforce road rules and give traffic insights without the need of physical interference. Common methods are background subtraction, motion detection, and/or convolutional neural network (CNN). The first two methods suffer from inability to differentiate classes and occlusion, whereas CNN suffer from computational complexity. A system based on You Only Look Once (YOLO) as detector and Kalman filter as tracker is proposed. TensorRT and frame sampling are used to further optimize the inference time. From experiment using BrnoCompSpeed dataset on i5 9600k and RTX 2060 Super environment, proposed system runs the entire process at 118 FPS with mean average error (MAE) of 0.96 km/h and [-3,2] error interval at 93.81%. Frame sampling can be used to further improve the FPS, with 1/5 sampling improves the speed by 50% to 177 FPS with only 0.11 km/h MAE tradeoff to 1.07 km/h.