HVD-Net: A Hybrid Vehicle Detection Network for Vision-Based Vehicle Tracking and Speed Estimation

Abstract

Worldwide, many urban areas have experienced rapid urbanization, population, and vehicle growth, compared to available road infrastructure. Among all categories of accidental deaths, the most fatalities occur because of traffic accidents. The likelihood, severity, and fatality of automobile collisions are all increased by speeding. Autonomous Speed Limit Violation Detection (SLVD) mechanism offers the best way to estimate speed with high precision using Intelligent Transportation Systems (ITSs) technology. Vision-based Vehicle Speed Monitoring (VSM) pipeline is proposed in this work, which includes mechanisms for vehicle detection, tracking, and speed measurement. VSM pipeline is based on a three-tier architecture, exploiting single RSU-camera. In the first tier, a real-time CNN-based Hybrid Vehicle Detection Network (HVD-Net) is designed for vehicle detection. The HVD-net utilizes multi-level and multi-scale features to minimize the impact of vehicle scale variation and maximize detection accuracy. Secondly, a Simple Online Real-time Tracker (SORT) along with HVD-Net is adopted to track vehicles' trajectories. The tracker utilizes a Kalman filter for vehicle state estimation and a Hungarian algorithm to solve the multi-vehicle association problem. Finally, a mechanism is presented to estimate the vehicle speed in the motion plane to mitigate the speed overestimation. Moreover, speed violation detection mechanism is also presented. Empirical evaluations on three datasets demonstrate that the proposed vehicle detection, tracking, and speed estimation schemes perform better when compared with the relevant and state-of-the-art schemes. The estimated pipeline vehicle speed is achieved with accuracy of 87.242%.