Urban Mobility Insights from CCTV: A Deep Learning Approach to Traffic Flow Monitoring

Abstract. This study explores the feasibility of using closed-circuit television (CCTV) data and computer vision methods, particularly YOLO (You Only Look Once) and ByteTrack, for traffic flow monitoring in Pavia, Iloilo. Two 24-hour videos from opposing lanes of a major road section were processed using YOLOv8 for vehicle detection and ByteTrack for multi-object tracking, using image rectification through homography and known vehicle measurements to estimate vehicle counts, speed, volume, and density. The models achieved high performance metric scores, with mAP50 values of 0.91 (GT Mall) and 0.89 (Robinsons). Results of the traffic flow analysis showed expected patterns: (a) vehicle speeds decreased as traffic volume and density increased, and (b) peak volumes and densities occurred during commuting hours. The interaction with and effects of between illumination and occlusion were considered during data preparation, resulting in mean absolute error rates below 3% compared with manual vehicle counts. By generating per-frame logs of traffic flow, the study shows the potential of computer vision–based monitoring systems serving as a supplement for evidence-based policymaking and urban planning for congestion management, road safety, and infrastructure planning at a local scale.