Video Dataset Labeling Using Active Learning with Applications in Vehicle Classification and Traffic Flow Rate Measurement

Intelligent Transportation Systems (ITS) offer a means to increase efficiency in road management, safety, and traffic enforcement. The Philippines, particularly Metro Manila, is notorious for its high levels of traffic congestion result in significant economic loss. It is possible to accumulate large amounts of traffic video data by installing traffic cameras but the manual preparation of such custom datasets for ITS applications is taxing and laborious. In this paper, we develop an algorithm for vehicle detection, classification, and flow rate measurement algorithm for ITS applications using YOLOv7 and a tracker trained on a custom dataset provided by the UP National Center for Transportation Studies (NCTS) augmented with active learning algorithm. Its detection, classification, and tracking performance were compared to that of a model trained without using active learning. The results indicate that using uncertainty-based active learning algorithm is effective in improving the model's tracking capability. The best results from the active learning models with the tracker was able to achieve a higher HOTA value of 67.423 vs 67.355 (+0.068%) for the first evaluation, and 71.652 vs 70.614 (1.038%) for the second evaluation on the NCTS tracking evaluation sets. For a specific sequence in DETRAC, the improvement is 69.842 vs 69.84 (+0.002%). At the third cycle of training, the active learning model counts better with a total count of 98 vs 100 from a true count of 91.