ESTIMATION OF VEHICLE GEOMETRY AND AXLE CONFIGURATION FROM TRAFFIC SURVEILLANCE CAMERA

Abstract

This paper proposes a method of above traffic parameter extraction from a stationary traffic surveillance camera. Combined with vanishing points from automated calibration method work by previous researchers, the proposed method utilized Yolov4 and morphology method for detection and the MedianFlow tracker for tracking 3 feature points from car wheels. Compared to the 3-dimensional bounding box method on ground truth dimension data with the help of a fine-grained classification model, our algorithm reduces mean absolute error for length estimation by 77%(1.14m to 0.26m). Then for the extraction of axle configuration on highway traffic, the proposed method applies DeepSORT tracking and YOLOv5-based wheel detector which is trained from 500 vehicle image data. This method prove to be with high accuracy on axle counting results with over 90% for each axle type and 93% accuracy on axle distance estimation.