Convolutional Neural Network for Object Detection System for Blind People

Abstract

Blind or visually impaired people are usually unaware of the danger that they are facing in their daily life. They may face many challenges when performing their daily activity even in familiar environments. This work proposed a smart object detection system based on Convolutional Neural Network (CNN) to provide a smart and safe living for visually impaired people. To reduce the complexity load, region proposals from the edge maps of each image were produced using edge box algorithm. Then, the proposals passed through a fine-tuned CaffeNet model. The object scene was captured by the webcam in real time and the feature of the image was extracted. After that, audio-based detector was generated on the detected object to notify the visually impaired people about the identified object. The result was evaluated by using mean average precision (mAP) and frame-per-second and it was found that the Single Shot MultiBox Detector (SSD) reduces the complexity and achieves higher accuracy as well as faster speed in object detection compared to Fast R-CNN.