AIoT-Based Smart Traffic Light Management System using YOLO Object Detection and Transfer Learning

Efficient traffic management is essential in urban areas having street congestions, as this has a great impact on people mobility, safety, and quality of life. In this paper, we propose a smart traffic management system leveraging artificial intelligence of things (AIoT) to enhance traffic flow. It utilizes real-time data of traffic densities at different intersections to dynamically control the timing of traffic light signals at these intersections. Live video feeds from street cameras are analyzed using the You-Only-Look-Once (YOLO) object detection algorithm combined with transfer learning to accurately count vehicles at intersections. This information is periodically uploaded to the smart city cloud, where it can be stored and analyzed further for making decisions and improving services including roads and traffic administration citywide. In this paper, we evaluate the latest two YOLO algorithms: YOL011 and YOL012. Three models per each YOLO version are used: nano ($n$), small (s), and medium (m). We train and validate the models on NVIDIA Tesla T4 GPU from Google Colab using the COC0128 dataset. Performance validation results show that YOL011n model is the fastest with inference time of 2.8 ms/image and speed of 357 fps, while the slowest model is YOL012m with inference time of 15.1 ms/image and speed of 66 fps. In addition, YOL011n is 1.93 times faster than YOL012n, which is similar in size. Moreover, we find that any YOL011 model in general has a mean average precision (mAP) similar to its YOL012 counterpart model and runs faster. The results also show that YOL011s and YOL012s have mAP similar to the medium-size models across Intersection-over-Union (IoU) thresholds range [0.50:0.95]. They achieve 0.84 and 0.87 mAP@[0.50:0.95], respectively, with significantly less GPU memory and computing resource requirements and much higher speeds, 2.51 times for YOL011s vs. YOL011m, and 1.78 times for YOL012s vs. YOL012m. We also do a comparison with previous related work.