This study addresses the critical challenge of improving autonomous navigation for visually impaired individuals by proposing a novel pedestrian lane detection model. Using a lightweight neural network based on vision transformers, the model integrates a multitasking learning architecture to detect traffic lights at intersections and tactile paving on sidewalks, while simultaneously regressing the walking direction in both environments. This dual approach ensures high accuracy while maintaining computational efficiency, making it suitable for deployment in IoT devices. The proposed model achieves a classification accuracy of 90.91% and an average path direction error of 6.19 \(^{\circ }\) , demonstrating its efficacy in real-world scenarios. By enabling precise and reliable navigation, this model offers a significant advancement in assistive technologies that empower vision-impaired users with enhanced mobility and safety.