Towards Efficient and Robust Night-time Vehicle Flow Monitoring via Lidar-based Detection

Abstract

The monitoring of vehicle flow is critical to enable a variety of intelligent transportation systems (ITSs). Traditional vehicle flow monitoring solutions are mainly based on roadside cameras, which may suffer serious performance deterioration in dark environments. In view of this, this paper proposes a Lidar-based vehicle flow monitoring system, which consists three parts: target detection module, vehicle flow counting module and vehicle counting visualization module. Specifically, the target detection module is built based on self-training data and the YOLOv4 network. Vehicle information is collected and preprocessed to speed up the target detection and enhance the accuracy. The vehicles and their positions are then obtained by performing inference with the trained weights for Lidar-based vehicle detection. On this basis, the vehicle counting module applies a multi-object tracking technique to monitor the vehicles which are nearby the detected one. Additionally, the Hungarian algorithm is used to match the surrounding vehicles. In vehicle counting visualization module, we visualize the system output through OpenCv. Finally, we build the system prototype and evaluate the algorithm performance in realistic environments under different night-time traffic situations. The experimental results demonstrate the practicability and robustness of the proposed solutions.