BlindAssist: Object Detection for the Visually Impaired using CBAM and Ghost Convolutions

Visually impaired individuals often face challenges in detecting and avoiding obstacles, compromising their mobility and safety. To address this, BlindAssist is proposed as an indoor object detection system that identifies objects in real time and provides directional audio cues. To enhance detection accuracy, three variations of the Rep-Net with Cross-Stage Partial and Efficient Layer Aggregation Network (RepNCSPELAN) module are introduced: with Ghost Convolutions, with Convolutional Block Attention Module (CBAM), and combining the two. These modules replace the original RepNCSPELAN in the YOLOv9s architecture. Among the variants, YOLOv9s with CBAM-RepNCSPELAN achieved the highest performance, improving mAP50 by 8.09% and mAP50-95 by 2.23%, with an inference speed of 13.36 ms per image. A paired t-test (p = 0.00026) confirms the statistical significance of the results. These findings highlight the potential of attention-enhanced lightweight architectures, particularly CBAM-RepNCSPELAN, in delivering accurate and efficient real-time object detection for assistive technologies supporting the visually impaired.