Environment Recognition from Spherical Camera Images Based on Multi-Attention DeepLab

Abstract

Electric wheelchair is an easy-to-operate means of transportation that does not require physical strength. With the number of electric wheelchair users increasing in recent years, the increase in traffic accidents becomes a problem. Therefore, by developing an autonomous electric wheelchair, it is expected that the risk of accidents will be reduced and the convenience of the electric wheelchair will be improved. Environment recognition is indispensable for the development of autonomous electric wheelchairs. We propose a semantic segmentation method for recognizing 16 objects in traffic environment. This paper examines the improvement of problems such as the high price of autonomous electric wheelchairs due to the increase in the number of sensors used, which has been a concern in related research. Therefore, we use panoramic images acquired by a spherical camera as input data, and extern the Multi-Attention Deep Lab algorithms fitting for the recognition of distorted images. A new CNN model is constructed using Deep Lab v3+, scSE Block, Pairwise Self-Attention, and Joint Pyramid Up-sampling. We conducted a recognition experiment using images taken on campus and verified its effectiveness. (Comparing to DeepLab v3+, IoU and Dice showed a 3.5% and 3.6% improvement in accuracy, respectively.)