As the aging population continues to grow, wheelchairs have emerged as crucial assistive tools in the elderly's everyday routines. Yet, the prevalent safety risks in traditional electric wheelchairs, mainly due to human operational errors, underscore the need for advancements. Addressing these concerns, this study introduces an innovative road segmentation approach utilizing an enhanced U-Net model tailored for intelligent wheelchairs. This technique adeptly segregates road surfaces, pedestrians, and obstacles, significantly bolstering the safety of wheelchair navigation. To curtail computational demands, the paper integrates a lean feature extraction mechanism inspired by GhostNet. Moreover, it presents a novel feature fusion strategy that marries coordinate attention mechanisms with skip connections, boosting the model's capacity for synthesizing information. The inclusion of pruning strategies effectively diminishes the model's parameter count, streamlining its efficiency. Empirical assessments reveal that our refined U-Net-based road segmentation method attains a mean Intersection over Union of 77.45% and a mean pixel accuracy of 85.62%, marking improvements of 4.58% and 4.82% over the traditional U-Net benchmarks, respectively. In real-world deployments within intelligent wheelchair systems, the proposed solution exhibits exceptional accuracy and robustness, heralding significant implications for enhancing mobility and safety for the elderly demographic.
|