Automated Dhaka City Vehicle Detection for Traffic Flow Analysis Using Deep Learning

Abstract

There are many ways to stop traffic jams from spreading, and one of the most effective is to detect the vehicle. The uniqueness of Dhaka's traffic situation creates a complicated and difficult occurrence, with over eight million passengers passing through the city every day in a 306 square kilometer area. To address this issue, our research includes a deep learning methodology for autonomous vehicle detection and localization from optical scans. Data preparation was done using annotated data from PoribohonBD with vehicle images. Vehicle detection is a critical step in the development of intelligent transportation systems (ITS). The challenges of vehicle detection on urban roads arise from the camera position, context variations, obstruction, multiple current frame objects, and transportation pose. The current study provides a synopsis of state-of-the-art vehicle detection techniques, which are classified thus according to motion and aesthetics techniques, beginning with frame differencing and background subtraction and progressing to feature extraction, a more complicated model in comparative analysis. The pre-processed data, as well as the fine-tuning hyper parameter, then input into the cutting-edge YOLOv5s deep learning model for autonomous vehicle detection and recognition. In the end, the training accuracy averaged 0.79% to detect vehicles in all classes.