Traffic Analysis Using Deep Learning and DeepSORT Algorithm

Abstract

With the help of advanced technology available, intelligent traffic has been implemented on roads to improve traffic congestion. Traffic analysis is essential to improve traffic flow in traffic light junctions. Therefore, this project analyzes vehicle speed, type, and count. The Jupyter notebook, Google Colaboratory, and Visual Studio Code are utilized to create live code. For multiple object tracking, Simple Online Real-time Tracking with Deep Association Metric (DeepSORT) algorithm is used with the help of the state-of-the-art object detection model You Only Look Once version 4 (YOLOv4). YOLOv4 is chosen as it is a simpler approach for object detection as compared to the regional proposal method, and it often takes far less processing time. Other than that, TensorFlow acts as an open-source platform for machine learning. It is a useful tool and library to provide workflows with high-level APIs. The average accuracy for each parameter is taken from the tabulated accuracy in different conditions: daytime, nighttime, and rainy weather. As a result, the system is able to detect vehicles with 83.65% average accuracy. There is a slight error in bad weather, nighttime, and traffic congestion for the vehicle count. Therefore, the average accuracy for vehicle count is 65.98%.