Vehicle Speed Estimation Modeling Using YOLOv8 and DeepSORT on Road Traffic Videos

Abstract

Vehicle speed monitoring is an important component of traffic analysis because excessive speed can increase traffic risk. This study develops a video-based vehicle speed estimation system using YOLOv8 for vehicle detection and DeepSORT for multi-object tracking. The system processes a public road-traffic video frame by frame, assigns a consistent tracking ID to each detected vehicle, determines its movement direction, and estimates speed from the travel time between two virtual lines. A pixel-to-meter conversion factor is applied to obtain the distance used in the speed calculation. The main model uses YOLOv8n with a confidence threshold of 0.45, while DeepSORT uses max-age 30, n-init 3, and max cosine distance 0.4. The results show that most YOLOv8-DeepSORT speed estimates are concentrated around 40–90 km/h, with an average estimated speed of 70.4 km/h. One extreme observation approaches 200 km/h and is treated as an outlier. As an additional baseline, SSD was tested on the same video. SSD produced 57 upward and 14 downward vehicle records, while its speed distribution was more dispersed, with an average of 91.3 km/h and several high-valued observations. Because the study does not use ground-truth speed or direct real-world distance calibration, the results are interpreted in terms of system consistency rather than absolute speed accuracy. Overall, YOLOv8 combined with DeepSORT is more suitable for the speed-estimation task evaluated in this study.

Keywords – vehicle speed estimation, YOLOv8, DeepSORT, object detection, object tracking, traffic video, SSD.