Abstract
Indoor object detection and recognition present one of the most crucial tasks for computer vision and robotic systems. Developing new intelligent autonomous robots is required in various applications including blind and visually impaired people assistance navigation and smart healthcare. Intelligent robots navigation is still a very challenging problem as it involves various aspects including indoor objects detection, recognition and scene understanding. We propose in this work to develop an indoor object detection system that can be used for intelligent vision of robotics applications. We ensure in this paper a lightweight implementation of the system using EfficientDet neural network. The proposed work presents a vision-based detection system able to work on real mobile robots by studying and considering their limited resources implementations. To ensure a lightweight implementation of the proposed indoor objects detection system and to design a deployable system in mobile robots application, we applied the weights pruning technique. To contribute for an embedded implementation of the proposed system, we used a pruning method which successfully reduced the network size, complexity and computation resources. Experimental results have demonstrated the robustness of the proposed indoor object detection system that can be deployed for indoor robotics assistance navigation systems. Based on the obtained results, we note that the proposed system achieved very competitive results in terms of detection precision as well as processing time. The proposed system can runs in low-end devices as we succeeded to reduce the parameters and FLOPs number, we achieved 89% on the testing set of the proposed indoor data set for EfficientDet D2. We achieved 31 FPS for the basic EfficientDet model and 38 FPS for the pruned model.