Abstract
Building new systems used for indoor objects detection and indoor assistance navigation presents a very crucial task especially in artificial intelligence and computer science fields. The number of blind and visually impaired persons (VIP) is increasing day by day. In order to help this category of persons, we propose to develop a new indoor object-detection system based on deep convolutional neural networks (DCNNs). The proposed system is developed based on the one-stage neural network RetinaNet. In order to train and evaluate the developed system, we propose to build a new indoor objects dataset which also presents 11,000 images containing 24 indoor landmark objects highly valuable for indoor assistance navigation. The proposed dataset provides a high intra and inter-class variation and various challenging conditions which aim to build a robust detection system for blind and visually impaired people (VIP) mobility. Experimental results prove the high detection performances of the developed indoor objects detection and recognition system. We obtained a detection accuracy reaching up to 98.75% mAP and 62 FPS as a detection speed.