Nowadays, urban traffic organization is becoming increasingly complex, and people have higher requirements for accurate and real-time road image processing. In this article, according to the characteristics of road images, it is very important to reasonably select the convolutional neural network structure and optimize it. In the optimization, the learning efficiency of the neural network was improved by adjusting the number, size and step size of the convolutional neural network, and selecting the excitation functions such as modified linear units, and the deviation of covariates was reduced by using batch standardization and other methods to speed up the convergence of the network. In this article, adaptive moment estimation, random gradient descent and other optimization algorithms were used to adjust the weights of neural network, so as to achieve the best image edge detection learning effect. In this article, a new method based on local least squares support vector machine was adopted, and it was improved by this method, so that it can better detect the edge of road images. The number of images in sunlight was 1000; the detection accuracy was 95.0%; the robustness score was 4.8. The number of images under shadow illumination was 800; the detection accuracy was 90.5%; the robustness score was 4.2. This article is beneficial to improve the edge detection accuracy of urban roads.