A Fast and Lightweight Network for Road Lines Detection Using Mobile-Net Architecture and Different Loss Functions

By using the line detection system, the relative position of the self-driving cars compared to other cars, the possibility of leaving the lane or an accident can be checked. In this paper, a fast and lightweight line detection approach for images taken from a camera installed in the windshield of cars is presented. Most of the existing methods consider the problem of line detection in the form of classification at the pixel level. These methods despite having high accuracy, suffer from two weaknesses of having the high computational cost and not paying attention to the general lines content information of the image (as a result, they cannot detect if there is an obstacle). The proposed method checks the presence of lines in each row by using the row-based selection method. Also, the use of Mobile-net architecture has led to good results with fewer learning parameters. The use of three different functions as cost functions, with different objectives, has resulted in obtaining excellent results and considering general content information along with local information. Experiments conducted on the TuSimple video image collection show the suitable performance of the proposed approach both in terms of efficiency and especially in terms of speed.