Abstract
Obstacle detection remains vital in autonomous navigation and assistive technologies, especially for visually impaired individuals. This work introduces an enhanced obstacle detection framework based on a Bilateral Vision Transformer and Convolution Kernel Neural Network (BViT-CKNN). The system incorporates stereo vision data and applies a bilateral filter to reduce noise while preserving edge details. A Vision Transformer (ViT) model is then used for global feature extraction, and a Convolution Kernel Neural Network (CKNN) captures fine-grained local features. Evaluated using the COCO dataset, the proposed BViT-CKNN achieves superior performance in precision (0.93), recall (0.91), F1-score (0.92), and Mean Absolute Error (MAE) reduction (3.16%) compared to existing methods.

This publication has 30 references indexed in Scilit: