Highway asset detection is a core technology in intelligent highway maintenance. However, traditional detection algorithms face issues such as high computational complexity and the misdetection or missed detection of small targets, making them unable to meet the demands for both accuracy and real-time performance. To ensure the optimal performance of highway infrastructure, developing efficient on-board highway asset detection algorithms is essential. In this study, we applied the k-means++ clustering algorithm to re-cluster the width and height of labeled target boxes in the training set, obtaining optimal prior box sizes and addressing the issue of target size diversity. For vehicle-mounted scenarios, we adopted a lightweight network architecture, replacing the CSPDarknet53 backbone of Yolov5 with MobileNetV3-large as the main feature extraction network. Additionally, to counteract the potential decline in detection performance due to the reduced complexity of the backbone network, we introduced an improved Local Normalization Attention Mechanism (L-NAM) module into the last convolutional layer of the neck network. This effectively mitigates false positives and false negatives for small targets.We propose a lightweight Yolov5s algorithm tailored for vehicle-mounted highway asset detection. Experimental results on a custom dataset show that the improved algorithm achieves an average precision of 98.2%, increases FPS to 91, and reduces the computational load in GFLOPs from 15.8 to 2.3. The proposed lightweight Yolov5s algorithm significantly reduces parameter count while maintaining high detection accuracy, providing an efficient and viable solution for vehicle-mounted highway asset detection.