Real-time Object Detection for Visually Impaired with Optimal Combination of Scores

In this paper, we propose a computer vision based object detection mechanism for visually impaired individuals by optimally combining the detection scores to aid their indoor navigation. Proposed framework comprises of three stages, viz. image acquisition, object detection and class estimation followed by score combination to generate the final classes of objects. The detection is performed on the image frames captured through wireless webcam and by applying two powerful deep learning algorithms, SSD and YOLO, separately. After generating the scores for each class, they are optimally combined using Proportional Conflict Resolving principle to generate fused set of scores. Based on these score values, object category is chosen with the category having the highest probabilistic score. The particular audio file narrating the object name is selected and the audio signal is sent to the blind user through the attached wearable headphone. The proposed framework achieved a high Mean Average Precision score of 86.15. Model performs well in an indoor environment and successfully aids a blind individual in mobility and mapping surrounding environment effectively.