Intelligent Pathway Perception System using YOLO and Multi-Architecture Deep Learning for Obstacle and Floor Hazard Detection

Independent and safe movement of the visually impaired people is a very urgent issue, especially in indoor areas, where the dynamic barriers and physical features like furniture, stairs and uneven surfaces are major hazards. The current assistive systems are usually concentrated on specific tasks, leading to a disjointed understanding of the environment and a lack of reliability. The proposed framework uses the YOLOv8 to detect obstacles in the indoor environment, then extracts features using parallel CNN and ResNet-50 and fuses them with an attention mechanism to get the surface of the pathway, which is helpful for context-aware perception. Training and evaluation are carried out on the selected categories of indoor objects from the Microsoft COCO dataset, derived from the task. The ResNet-50 model is initialized with ImageNet-pretrained weights, while pseudo-labels are introduced for surfaces to support surface inference by leveraging the COCO object-scene relationships in a weakly supervised manner. Therefore, it does not need an extra floor-hazard training data set, however, the feature initialization is based on the principle of transfer learning. The results of the experiments prove that APPF has a detection accuracy of 96.8%, a precision of 95.1% and, a recall of 94.3% and a mAP at 0.5 of 0.91 in object detection tasks. The context-aware surface inference module has an accuracy of classification of 95.5% and a false detection rate of 4.5%. Relative analysis with baseline CNN and single-stage detectors demonstrates 17.4% improvement in the detection performance and 36.5% decrease in the latency, maintaining real-time inference at 27 FPS. It is innovative in that it uses only the COCO dataset, but it allows dual-level perception to be done by architectural innovation and contextual learning to indoor settings.