Traffic sign installation angle measurement method based on the You Only Look Once algorithm and binocular stereo vision

We propose a framework for traffic sign installation angle measurement. The framework integrates an enhanced You Only Look Once-Lite (YOLO-Lite) detection model with binocular stereo vision. The enhanced YOLO-Lite detection model incorporates a lightweight attention convolution (LAC) module and adaptive downsampling mechanism (ADM). Both the LAC module and ADM contribute to improving the accuracy of traffic sign detection. The LAC module enhances traffic sign detection accuracy by integrating spatial and channel attention mechanisms, enabling the extraction of discriminative features from small targets in complex backgrounds. Also, the LAC module reduces the number of training parameters through depthwise separable convolutions, contributing to improved computational efficiency without compromising detection accuracy. Simultaneously, the ADM module mitigates information loss during multi-scale processing through adaptive downsampling strategies that preserve critical features. An Iterative Geometry Encoding Volume with MobileNetV2 enhances performance through MobileNetV2-based lightweight encoding and single-iteration convolutional gated recurrent unit optimization. The model maintains the consistency of contour matching and multi-scale depth estimation while reducing the computational burden. The parameters and inference time of the overall framework are only respectively 7.8 M and 1.37 s, with an angle measurement error of 4.80%. This meets the accuracy and real-time requirements for traffic sign installation angle measurement implemented on embedded devices. The proposed framework provides an efficient, cost-effective solution for the standardized installation and maintenance of road traffic infrastructure.