Environment Recognition from A Spherical Camera Image Based on DeepLab v3+

Abstract

The number of users of electric wheelchairs has been increasing in recent years because it is easy to operate the electric wheelchair and do not require physical strength. However, the traffic accidents are also increasing because of the large number of wheelchairs. The development of autonomous electric wheelchairs is expected to reduce the risk of accidents and improve the convenience of electric wheelchairs. Environmental recognition is essential for the development of autonomous electric wheelchairs. In this paper, we propose a method for recognizing roads, sidewalks, buildings, electric wheelchair drivers, poles, electric wheelchairs, vegetation, curbs, sky, pedestrians, lanes, cars, steps, and bicycles. For recognizing those objects, we use a panoramic image acquired from a spherical camera. As the machine techniques, we use DeepLab v3+, a semantic segmentation algorithm based on Convolutional Neural Network (CNN). In the proposed method, a new CNN model is constructed by adding deformable convolution, SE-block, and MobileNet v2 to DeepLab v3+ into the original DeepLab v3+. In the experiment, IoU 38.8% and Dice of 46.7% were obtained.