Tiny-PIRATE: A Tiny Model With Parallelized Intelligence for Real-Time Analysis as a Traffic countEr
Due to the rapid growth in the number of vehicles over the last decade, there has been a dramatic increase in demand for highway capacity analysis. Vehicle counting, in particular, has become a key element of vision-based intelligent traffic systems deployed across metropolitan areas. Most methods solved the vehicle counting problem under the assumption of state-of-the-art computing systems. However, large-scale deployment of such systems for multi-camera processing is very inefficient. With the recent advancement of cost-efficient Internet-of-Things (IoT) devices alongside machine learning methods developed specifically for such devices, solving the vehicle counting problem for real-time traffic analysis on IoT edge devices, and thereby facilitating its large-scale deployment have become highly favorable. In this paper, we propose a framework of vehicle counting designed specifically for IoT edge computers which follows the detection-tracking-counting (DTC) model. The proposed solution aims at addressing the multimodality of contextual dynamics in traffic scenes with a small detector model, a robust tracker and a counting process that accurately estimate both a vehicle's motion of interest and its exit time from observation areas. Experimental results on AI City 2021 Track-1 Dataset showed that ours outperformed related methods with promising results regarding both accuracy and execution speed.