A Smart Obstacle Avoiding Technology Based on Depth Camera for Blind and Visually Impaired People

Jian He,Xuena Song,Yuhan Su,Zhonghua Xiao
DOI: https://doi.org/10.1007/s42486-023-00136-7
2023-01-01
CCF Transactions on Pervasive Computing and Interaction
Abstract:It remains challenging to assist BVI individuals in outdoor travel nowadays.In this paper, We propose a set of low-cost wearable obstacle avoidance devices and introduce an obstacle detection algorithm called L-PointPillars, which is based on point cloud data and is suitable for edge devices. We first model the obstacles faced by BVI individuals during outdoor travel and then establish a mapping between the information space and physical space based on point clouds. We then introduce depthwise separable convolution and attention mechanisms to develop L-PointPillars, a fast neural network for obstacle detection. This network is specifically designed for creating wearable obstacle detection devices. Finally, we implemented a wearable electronic travel aid device (WETAD) based on L-PointPillars on the Jetson Xavier NX. Experiments show that while L-PointPillars reduces the number of parameters in the original PointPillars by 75%, WETAD achieves an average obstacle detection accuracy of 95.3%. It takes an average of 144 milliseconds to process each frame during outdoor travel for BVI individuals, which is more than twice as fast as the Second network and 31% improvement compared to PointPillars.
What problem does this paper attempt to address?