With the rapid development of intelligent transportation systems

autonomous driving technologies have imposed higher requirements on the real-time performance and accuracy of road detection. However

traditional deep learning models often suffer from high computational complexity and large storage demands in road detection tasks

limiting their application on in-vehicle and edge computing devices. This paper focuses on the application of model lightweighting techniques in road detection

aiming to significantly reduce the computational complexity and storage requirements while maintaining high detection accuracy through model structure optimization

pruning

and quantization.This study first reviews the current status of model lightweighting technologies in road detection both domestically and internationally

analyzing the advantages of lightweight model structures such as MobileNet and the development of pruning and quantization techniques. Subsequently

a lightweight road detection model architecture based on an improved MobileNet is proposed. By optimizing network layer structures

adjusting convolution kernel sizes and numbers

and reducing redundant computations

the model dynamically adjusts computational load during runtime through sparse networks and dynamic pruning structures. Effective evaluation metrics and optimization algorithms are also designed to address the critical balance between accuracy and lightweighting.In the experimental section

comparative experiments were conducted on the YOLOv11 model using existing lightweighting techniques. Results show that the lightweight model retains high detection accuracy while significantly improving inference speed and reducing model size

making it more adaptable to various road scenarios and hardware conditions. The findings not only provide an efficient model solution for road detection tasks but also offer insights for applying lightweighting techniques to other computer vision tasks.