Indexed by:
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. © 2023, China Computer Federation (CCF).
Keyword:
Reprint Author's Address:
Email:
Source :
CCF Transactions on Pervasive Computing and Interaction
ISSN: 2524-521X
Year: 2023
Issue: 4
Volume: 5
Page: 382-395
Cited Count:
SCOPUS Cited Count:
ESI Highly Cited Papers on the List: 0 Unfold All
WanFang Cited Count:
Chinese Cited Count:
30 Days PV: 2
Affiliated Colleges: