Road transport is critical to economic development, yet pavement defects like cracks and potholes pose significant safety risks. Traditional detection methods are inefficient and resource-intensive, while existing deep learning models often fail to meet real-time requirements on edge devices. To address these challenges, we propose LPDDN (Lightweight Pavement Defect Detection Network), a novel embedded real-time detection model. LPDDN integrates a Lightweight Information Sharing Detection Head (LISD Head) to reduce model complexity, a WIoU loss function to enhance detection accuracy, and a Layer-Adaptive Magnitude-based Pruning (LAMP) method to optimize computational efficiency. Experiments on the China-MotorBike dataset demonstrate LPDDN’s superior performance, achieving 94.9% mean average precision (mAP) with only 858,941 parameters. Deployment on NVIDIA Jetson Xavier NX further proves its real-time capability with detection times under 50 ms. LPDDN offers an efficient solution for real-time pavement defect detection, balancing accuracy and computational efficiency for practical applications.