CNN Based Indoor Space Recognition System to Aid Visually Impaired People

Abstract

About 17,000 people in HK (2.4% of the total population) are visually impaired. Besides guide dogs and sticks, the indoor navigation relays on high-cost appendant, such as sensors, RFID and Wi-Fi. Convolutional Neural Network (CNN) has achieved impressive performance in object detection and scenes recognition. However, it is rarely applied in the field of indoor navigation for the visually impaired. This project intends to propose a CNN based indoor scenes recognition system to improve mobility for the visually impaired. 6,100 images from Hong Kong Blind Union and online image dataset were collected to form the dataset. A CNN model based on AlexNet was optimized on this dataset and deployed in an android app. The resulted model achieves 85% accuracy in test dataset. Furthermore, the model’s behavior was studied through Class Activating Mapping technique, indicating how the model classify the scenes based on discriminative image regions and spatial features. This study has proved that using CNN is a potential solution to recognize the indoor scenes, which could be applied to give visually impaired people real-time navigation. Expansion of training dataset and combination with wearable devices could be done to further improve the performance.