Visual-LiDAR Place Recognition based on Evolutionary Game Theory and Neural Network

The deep development of place descriptors have made place recognition less challenging. However, besides the discrimination ability, the distance threshold is another determining factor for the place recognition online performance, which is less studied. In this paper, place recognition is regarded as a decision-making task, and a decision-making method based on neural networks is proposed. Firstly, considering a place recognition task with visual and LiDAR sensors, we construct a multiple information sources competition model based on evolutionary game theory. Based on this model, the key factors that are highly related to the final place recognition decision can be obtained. Then, we design and apply a neural network to learn the decision-making function. To balance the negative-positive place recognition samples, an acceptance-rejection sampling-based negative independent samples reduction method is proposed to refine the network training performance. The proposed method can make the robot get rid of using the preset descriptor distance threshold and also serve as a novel multi-modal fusion framework. Extensive experiments are conducted on public and our dataset collected by our mobile platform. The results demonstrate that our method can effectively improve place recognition performance in different environments. We will make our source code open to benefit the robotic community.