Integration of 2-D Sliding DFT Kernels and CNN for Edge Feature Extraction and Lane Detection
Nagineni Sukumar, P. Sumathi · IEEE Sensors Letters · 2025
An integrated vision-based system is proposed for lane detection which is built with sliding discrete Fourier transform (SDFT) kernels based cellular neural network (CNN) for edge and lane features extraction. RANdom SAmple Consensus (RANSAC) method is used for model fitting of lanes in the proposed lane detection system. The primary challenges addressed include high noise levels, variable lighting conditions, and complex shadowing on road surfaces. A novel edge detection kernel obtained from the 2-D SDFT basis function is described and evaluated on road images. The proposed edge detection kernels are integrated with CNN model as control templates and the conventional feedback template is retained as such. These modifications on CNN model improves the edge detection capability significantly under noisy conditions for the signal to noise ratio up to 10 dB. The performance evaluation of proposed edge detection method in terms of average gradient, entropy, and Pratt's figure of merit (PFOM) shows better performance than conventional methods. The lane detection system is implemented in real-time processor of compact reconfigurable input output (cRIO) system. The proposed system offers promising results in the following test environments: day time, rainy, back light, tunnel, and nighttime. The proposed lane detection is enhancing the reliability of advanced driver assistance systems (ADAS).