Training encoder-attention through fully-connected CRFs for efficient end-to-end lane detection model

Abstract

In this paper, we propose a novel training method to correct encoder-attention in semantic segmentation through fully-connected conditional random fields(CRFs) and train the attention of the encoder from the corrected encoder-attention. The novel training method allows the encoder to extract lane features robustly to the surrounding environment, and assists in interpolation of lane segments when the decoder upsamples from the features extracted by the improved encoder. Through the proposed training method, we present an efficient end-to-end lane detection model with few parameters and outperforming current state-of-the-art lane detection system. Deep learning models with various of structures and operations are being studied to improve the performance of the lane detection models for self-driving car around the world. However, because lane detection technology is an essential and important technology for self-driving technology, once a lane detection model is selected and applied to an embedded system, it is difficult to replace the model with another model whenever a performance improvement is needed. Therefore, in order to efficiently improve lane detection performance, instead of changing combinations of deep learning structures and operations, it is important to analyze and understand the internal structure and data flow of the lane detection model. In addition, it is also essential to find the factors that degrade the performance of the lane detection model. In this paper, we design a lane detection model using the encoder and decoder block of LinkNet structure which shows excellent performance in real-time semantic segmentation models. Besides, we analyze the physical meaning of encoder and decoder for lane detection. This resulted in that during upsampling of the decoder, noise of the surrounding background was found caused by passing to the decoder from wrong feature extraction of the encoder. Therefore, this paper proposes a novel training method to improve encoder performance through encoder attention and fully-connected CRFs. The proposed method is as follows. (1) One channel feature map representing an encoder is derived through encoder-attention. (2) The feature map of the encoder-attention is corrected considering the color information of the image and the binary information of the labeled lane through the fully-connected CRFs. (3) Encoder-attention is trained from target of the corrected feature map by fully-connected CRFs to learn post-processing of fully-connected CRFs. The proposed method extends receptive field of encoder and removes noise from background spatial features, instead of lane features. In this paper, to verify the effectiveness of the proposed method, we use three backbone networks to identify the difference of performance with or without the proposed method quantitatively. Through the proposed method, we present an efficient end-to-end lane detection model outperforming the performance of current state-of-the-art ENet-SAD and Spatial CNN. In addition, the performance of the model is verified with Tusimple dataset.