Autonomous Indoor Robot Navigation via Siamese Deep Convolutional Neural Network

The vast majority of indoor navigation algorithms either rely on manual scene augmentation and labelling or exploit multi-sensor fusion techniques in achieving simultaneous localization and mapping (SLAM), leading to high computational costs, hardware complexities and robustness deficiencies. This paper proposes an efficient and robust deep learning-based indoor navigation framework for robots. Firstly, we put forward an end-to-end trainable siamese deep convolutional neural network (DCNN) which decomposes navigation into orientation and localization in one branch, while achieving semantic scene mapping in another. In mitigating the computational costs associated with DCNNs, the proposed model design shares a significant amount of convolutional operations between the two branches, streamlining the model and optimizing for efficiency in terms of memory and inference latency. Secondly, a transfer learning regime is explored in demonstrating how such siamese DCNNs can be efficiently trained for high convergence rates without extensive manual dataset labelling. The resulting siamese framework combines semantic scene understanding with orientation estimation towards predicting collision-free and optimal navigation paths. Experimental results demonstrate that the proposed framework achieves accurate and efficient navigation and outperforms existing "navigation-by-classification" variants.