Vehicle tracking and classification for intelligent transportation systems using YOLOv5 and modified deep SORT with HRNN

Abstract

The increasing number of vehicles, traffic congestion, and security concerns have made it imperative to provide Intelligent Transportation Systems (ITS) with a potential system for vehicle surveillance, traffic monitoring, and vehicle control. A popular method for locating moving objects in videos is image subtraction, although it is not very effective because of its sensitivity to brightness changes. To achieve the best performance for vehicle tracking in traffic videos with dynamic lighting conditions, varying backdrops, and noises, a method that integrates image and video processing techniques is proposed in this paper. We utilized three datasets for validation of our intelligent vehicle detection system, covering daytime and nighttime conditions for unbiased evaluation. The frames were captured from the input traffic videos that represent the road in different traffic situations. The proposed methodology is categorized into three major stages: preprocessing, segmentation, and vehicle classification. In this work, we utilized color preprocessing to enhance vehicle classification and counting in an intelligent transportation system. To further reduce noise from the video frames and improve data quality, we utilized non-local means and a trilateral filter (NLMTF), improving edge preservation and contrast under low-light conditions. In the next stage, modified YOLOv5 is used for effective vehicle detection. By utilizing the advanced capabilities of YOLOv5, the system enhances the accuracy and effectiveness in detecting and localizing vehicles within the processed frames, ensuring precise vehicle presence capturing. Finally, we utilized the Modified Deep SORT Algorithm with a hierarchical recurrent neural network (HRNN) to track, count, and classify the vehicles. We achieved an accuracy of 94.75%, precision of 0.94, and recall of 0.92, compared to the ANN, MLP, CNN, and RNN, demonstrating the effectiveness of our proposed methodology.