A Transfer Learning Approach for Indoor Object Identification

Accessing new indoor environments is a well-known challenge for blind and visually impaired persons (VIP). In this work, we propose a new computer vision-based indoor object recognition system used especially for indoor wayfinding and indoor assistance navigation. The proposed recognition system is based on transfer learning techniques. This system is able to detect with a big performance three categories of indoor classes (door, stairs and sign). We developed an efficient and a robust indoor landmark identification system based on a lightweight deep convolutional neural network (DCNN). The proposed detection system is generic and performant enough to handle the large intra-class variation provided in the training and the testing sets. Experimental results have shown the big efficiency of the obtained systems by achieving high recognition rates.