关键词: 视觉同步定位与建图, 回环检测, 深度学习, 卷积神经网络加速, 移动机器人

Abstract: The deep learning based loop closure detection algorithm has been verified to be superior to traditional methods. However, the computation burden of deep learning is heavy, so it is often difficult to deploy large convolutional neural networks on mobile robots, while small convolutional neural networks perform poorly in large scenes. Therefore, this paper proposes a scheme to deploy large convolutional neural networks on mobile robots. Firstly, the feature graph is transformed into the feature vector by using the mixed global pooling layer. Experiments show that the performance of this method is equivalent to that of other more complex methods and the calculation is simpler. Then, a block-based floating-point convolutional neural network acceleration engine is proposed, which significantly reduces the computational energy consumption and causes almost no performance loss without retraining.

Key words: Key words: visual SLAM, loop closure, deep learning, CNN accelerator, mobile robot