Publication

Abstract : The large push for smart transportation and self-driving cars has brought out an array of research problems and a pertinent one among them is lane detection. With Convolutional Neural Network (CNN) coming up as the leading algorithm for this purpose, a lot of research has gone into the optimization of its performance. This paper proposes the application of Dynamic Mode Decomposition (DMD) algorithm for reducing the complexity of CNN for lane detection. The work involves implementation of a CNN model for lane detection and optimization of its inference time by reducing the number of layers while maintaining accuracy of the model with the help of DMD. The model implementation was done using TensorFlow library and the optimized model achieved accuracy levels similar to that of the baseline model. Both models were tested and it was found that the proposed model performed faster than the baseline; due to the lesser number of computations it performs as a result of reduction in its layers. This work shows that DMD as a background foreground separation tool, can be used to optimize the layers in a deep neural network trained to identify sparse details in images.