基于轻量化CNN的智能驾驶道路路面坑洼检测

Pothole Detection on Road Surfaces for Intelligent Driving Based on Lightweight CNN

智能驾驶过程中对道路路面坑洼检测能够提高车乘人员的舒适感。轻量化CNN通过LightSampleModule双分支结构使得计算量大大减少;卷积神经网络节点优化能够有效捕捉坑洼细节信息,减少有用特征的丢失;增加通道注意力机制获得不同通道的相关性,分组卷积(GroupConvolution)内划分若干个小块,每个小块使用不同的卷积核。实验仿真显示轻量化CNN算法检测准确率曲线高于CNN算法,损失值曲线低于CNN算法,召回率、平均F1分数相比CNN算法分别提高2.14%、2.08%。

Pothole detection on road surfaces during intelligent driving can enhance the comfort of passengers.The lightweight CNN significantly reduces computational load through a dual-branch structure in the Light Sample Module.Optimization of convolutional neural network nodes effectively captures detailed information about potholes and reduces the loss of useful features.The addition of a channel attention mechanism captures the relevance of different channels,while Group Convolution divides the network into several small blocks,each using different convolution kernels.Experimental simulations show that the accuracy curve of the lightweight CNN algorithm is higher than that of the CNN algorithm,the loss value curve is lower,and the recall rate and average F1 score are improved by 2.14%and 2.08%,respectively,compared to the CNN algorithm.