Camera-Based System for the Automatic Detection of Vehicle Axle Count and Speed Using Convolutional Neural Networks

Abstract

This paper outlines the development of a non-intrusive alternative to current intelligent transportation systems using road-side video cameras. The use of video to determine the axle count and speed of vehicles traveling on major roads was investigated. Two instances of a convolutional neural network, YOLOv3, were trained to perform object detection for the purposes of axle detection and speed measurement, achieving accuracies of 95% and 98% mAP respectively. Outputs from the axle detection were processed to produce axle counts for each vehicle with 93% accuracy across all vehicles where all axles are visible. A simple Kalman filter was used to track the vehicles across the video frame, which worked well but struggled with longer periods of occlusion. The camera was calibrated for speed measurement using road markings in place of a reference object. The calibration method proved to be accurate, however, a constant error was introduced if the road markings were not consistent with the government specifications. The average vehicle speeds calculated were within the expected range. Both models achieved real-time speed performance.