DrishT: A Dual-Stream Mobile Architecture for Real-Time Scene Understanding in Assistive Navigation

Autonomous navigation for visually impaired individuals requires simultaneous interpretation of spatial objects and contextual text—capabilities that existing systems address in isolation or require expensive dedicated hardware. We present DrishT , a mobile-optimized dual-stream architecture combining Single Shot MultiBox Detector (SSD) with VGG16 backbone for object detection and Convolutional Recurrent Neural Network (CRNN) for scene text recognition. Our system achieves 41.2% mAP@0.5 for object detection and 91.5% word-level accuracy for text recognition with 182ms end-to-end latency on mid-range Android devices (Snapdragon 660+). The key contribution is a semantic fusion layer that establishes spatial-textual correspondences through distance-weighted attention, achieving 84.3% accuracy in associating detected objects with embedded text labels. Through systematic optimization using TensorFlow Lite INT8 quantization, we reduce the combined model size to 14.5 MB (76% reduction from FP32) with only 1.8% mAP degradation. Evaluation on 1,000 real-world navigation images demonstrates practical utility for scenarios including menu scanning (92% text accuracy), sign reading, and obstacle identification. Our implementation addresses the cost barrier to assistive technology by enabling deployment on devices costing under $200, compared to proprietary solutions exceeding $3,500. Keywords: Assistive Technology, Computer Vision, Object Detection, Optical Character Recognition, Mobile Deep Learning, Real-time Systems, Accessibility