To address the challenges of accelerated deterioration of concrete bridges caused by overloaded vehicles, this paper proposes a multi-stage heterogeneous visual framework for overloaded vehicle identification. First, a block-wise foreground–background separation method based on two-dimensional correlation coefficients is introduced and integrated with an improved Gaussian Mixture Model (GMM) to achieve dynamic background modeling and robust foreground extraction from images. Next, the Fuzzy C-Means (FCM) clustering algorithm is employed to automatically localize vehicle regions. Subsequently, Histogram of Oriented Gradients (HOG) features of vehicle candidate regions, reduced by Principal Component Analysis (PCA), are extracted and combined with a Support Vector Machine (SVM) to eliminate non-vehicle objects. Finally, an enhanced YOLOv8 model is constructed for axle-count-based overloaded vehicle detection, in which Inception modules are embedded into the CSP Darknet backbone to capture multi-scale deep hierarchical features. Meanwhile, Canny edge detection and affine transformation are fused to optimize axle-counting recognition, and overloaded vehicles are classified in accordance with the Chinese national standard GB1589-2016. Experimental results on real-world concrete bridge surveillance scenarios show that the proposed method can significantly suppress noise in vehicle foreground extraction. After SVM post-processing, the vehicle purification accuracy reaches 98.75%, with a precision of 100% for the non-vehicle category. Compared with the vanilla YOLOv8, the proposed multi-stage heterogeneous visual framework improves the precision, recall, and mAP@50 by 8%, 12.5%, and 7.2%, respectively, for heavy-duty vehicle axle recognition. The axle-feature-based heavy vehicle recognition method achieves an overall identification accuracy of 92%.
Li et al. (Thu,) studied this question.