Abstract
Reliable object detection is crucial for autonomous driving, particularly in highway construction zones where early hazard recognition ensures safety. This paper introduces an enhanced YOLOv8s-based detection system incorporating a motion-temporal attention module (MTAM) to improve robustness under high-speed and dynamic conditions. The proposed architecture integrates a cross-stage partial (CSP) backbone, feature pyramid network-path aggregation network (FPN-PAN) feature fusion, and advanced loss functions to achieve high accuracy and temporal consistency. MTAM leverages temporal convolutions and attention mechanisms to capture motion cues, enabling effective detection of blurred or partially occluded objects. A custom dataset of 34,240 images, expanded through extensive data augmentation and 9-Mosaic transformations, is used for training. Experimental results demonstrate strong performance with mAP(IoU[0.5]) of 90.77 ± 0.68% and mAP(IoU[0.5:0.95]) of 70.20 ± 0.33%. Real-world highway tests confirm recognition rates of 96% for construction vehicles, 92% for roadside warning signs, and 84% for flag bearers. The results validate the framework's suitability for real-time deployment in intelligent transportation systems.