ALCDNet: Loop Closure Detection Based on Acoustic Echoes

Abstract

Loop closure detection is a critical component of simultaneous localization and mapping (SLAM) systems, essential for mitigating the drift that accumulates over time. Traditional approaches utilizing light detection and ranging (LiDAR) and cameras have been developed to address this challenge. However, these methods can be ineffective when there is a lack of visual cues, such as smoke, poor lighting conditions, and textureless environments. In this letter, we propose an efficient loop closure detection method that employs a speaker and microphone array to gather spatial structure information. First, our method uses a microphone array to capture echoes from finely designed signals emitted by the speaker. Second, we apply momentum contrastive learning (MoCo) to train an echo feature encoder to learn the implicit spatial features embedded in the echo signals. Finally, loop closure detection is performed by computing the cosine similarity of features output by the encoding network from echo information at different locations. Experiments conducted in typical indoor environments demonstrate that our method outperforms vision-based methods in most cases and can still achieve accurate loop closure detection in smoky environments where both LiDAR and vision-based methods fail. This makes it a viable and cost-effective complementary solution in environments with sparse texture features, unstable lighting conditions or smoke.