Research On Road Scene Understanding Algorithm

Vehicles equipped with autonomous driving systems or advanced driving assistance systems, i.e., Intelligent Vehicles come into our life gradually. Road scene understanding plays an important role for them, i.e., perceiving the road environment. Road scene understanding performs well in the normal weath-er or illumination conditions. However, the performance is decreased sharply under drastic weather (rainy day, snowy day, foggy day, etc.) or illumination(daytime, nighttime, etc.) variances. Improving the accuracy and robustness of the road scene understanding system under these challenging scenarios is badly needed. Based on these, we propose the vision-based cross-domain road scene understanding.Specifically, our proposed method consists of mainly two steps: a) cross-domain road scene retrieval, and b) cross-domain road scene dense correspon-dence. Given one road scene image in the challenging conditions, we first need to find the corresponding one in the normal conditions, i.e., cross-domain scene retrieval. The aim of cross-domain dense correspondence is to establish the dense relationships between road scene image in the challenging condition and the corresponding one in the normal condition. To be more specific, it refer-s to find the corresponding pixel or pixel region of one image in the normal condition when given a pixel or pixel region of the same location image in the challenging condition. The information of images in the normal conditions can be transferred to the images in the challenging conditions on the basis of the established dense correspondences.We summarize contributions of our paper below:·We propose a dense correspondence based transfer learning framework for perceiving the road environment for intelligent vehicles under dras-tic weather or illumination changes. To the best of our knowledge, this is the first time that the road scene understanding is approached by this framework.·We propose the feature transformation based on metric learning and sub-space alignment to improve the performance of the deep feature of the road scene.·We propose the cross-domain road scene dense correspondence algorith-m. In consideration of different data distribution among different weather and illumination conditions, we build the Cross-domain Deep Dense Cor-respondence Network (CDDCN).·To verify the proposed cross-domain road scene understanding frame-work, we carefully annotate samples to build cross-domain road scene data set, which consists of six different weather or illumination condi-tions, i.e., sunny day, cloudy day, snowy day, foggy day, night and rainy night. We collect pair-wise outdoor scene images across different weath-er or illumination changes from 30,000 cameras for pre-training the CD-DCN. Meanwhile, we collect 12,387 pair-wise road scene images from the same road route under different weather or illumination conditions for fine-tuning the pre-trained CDDCN.