Toward the Resilient SLAM in Adverse Illumination Environments with Learning-based Image Transformation

Abstract

Numerous bio-inspired vision robots are applied to various environments, which are usually in adverse illumination (low, intense, unstable light), bringing challenges to visual SLAM. Thus, it is of great importance to improve its robustness to sustain the autonomy of intelligent robots. However, most existing works focus on single and simple tasks rather than the SLAM explicitly, resulting in the lack of resilient SLAM towards challenging lighting. To break through the dilemma, we propose an illumination enhancement neural network tailored for visual SLAM. First, we introduce the optical flow constraints and LSTM (Long Short Term Memory) to exploit the potential consistency of the frame-wise movement and illumination in tempo-spatial sequences. Second, we implement the network with DSC (Depthwise Separable Convolution) and Conv-LSTM for a more lightweight network and less computation, considering the requirements for efficiency in SLAM. Finally, we integrate the trained network to ORB-SLAM2 in a modular way and develop a prototype system. Extensive experiments on public (KITTI and EuRoC), synthesized, and simulated sequences demonstrate the feasibility and superiority of our method with high accuracy and efficiency. To our best knowledge, this is the first SLAM integrated with a tailored lightweight enhancement network for adverse illumination.