Sensors, Journal Year: 2025, Volume and Issue: 25(6), P. 1839 - 1839

Published: March 15, 2025

Accurate obstacle recognition and avoidance are critical for ensuring the safety operational efficiency of autonomous robots in dynamic complex environments. Despite significant advances deep-learning techniques these areas, their adaptability environments remains a challenge. To address challenges, we propose an improved Transformer-based architecture, GAS-H-Trans. This approach uses grouped aggregation strategy to improve robot’s semantic understanding environment enhance accuracy its strategy. method employs dual-coupling optimize feature extraction global representation, allowing model capture both local long-range dependencies. The Harris hawk optimization (HHO) algorithm is used hyperparameter tuning, further improving performance. A key innovation applying GAS-H-Trans tasks implementation secondary precise image segmentation By placing observation points near obstacles, this refines recognition, thus flexibility motion planning. particle swarm (PSO) incorporated attractive repulsive gain coefficients artificial potential field (APF) methods. mitigates minima issues enhances stability avoidance. Comprehensive experiments conducted using multiple publicly available datasets Unity3D virtual robot environment. results show that significantly outperforms existing baseline models tasks, achieving highest mIoU (85.2%). In + PSO-optimized APF framework achieves impressive success rate 93.6%. These demonstrate proposed provides superior performance planning, offering promising solution real-world navigation applications.

Language: Английский