Simple and Effective Visual Place Recognition via Spiking Neural Networks and Deep Information

Visual place recognition (VPR) is a hot topic in current research on unmanned intelligent systems. However, the false recognition caused by changes in environmental conditions (e.g. weather, lighting) and resource limitations on the mobile robot are the main problems affecting the accuracy of visual place recognition, and how to use small robots to perform visual place recognition tasks in changing environments in an efficient and energy-efficient manner has important research and practical value. In this paper, we propose a simple and effective visual place recognition system called D-VPR, which simulates a bionic approach to visual place recognition. D-VPR makes robot look around and incorporates environmental depth information, the essence of this system is to optimize the weight response of neurons by processing the environmental information, and guide the neural network to carry out correct place recognition. We design and deploy the scheme on a mobile robot, and test it with our dataset, which is called PKU-Paradise. The results shows that the system has better robustness which is maximum at 61% to the interference of light changes, while having higher accuracy and lower power consumption than advanced visual place recognition methods. The implementation of this system makes a strong contribution to build a full-spiking neural network SLAM system in the future.