Abstract
Point cloud place recognition (PCPR) is a fundamental task in robotics and computer vision in the fields of autonomous driving, intelligent transportation, and augmented reality. To cope with the dynamic changes in scenarios and sensor types, PCPR models need to incrementally acquire, update, and accumulate knowledge for continuous evolution—an ability known as continual learning (CL). However, due to the dynamic distributions of incrementally acquired point cloud data, PCPR models often forget previous knowledge when acquiring new knowledge, i.e. catastrophic forgetting. To address this issue, this study proposes a novel CL method tailored for PCPR, which effectively extracts and fuses knowledge learned by the model across sequential point cloud data. First, a replay sample selection method is proposed, dynamically allocating a replay sample size to each training set based on information quantity and selecting replay samples based on spatial distribution. Second, a new CL framework composed of a prompt module and the two-stage training strategy is proposed, and the domain-specific knowledge captured from each training set by the prompt module is used for guiding the backbone network to extract features adapted to individual samples. Comprehensive experiments on large-scale public and self-collected datasets are conducted to validate the effectiveness of the proposed method. Compared with state-of-the-art (SOTA) methods, our method achieves 6.50% improvement in 𝑚𝐼𝑅@1, 7.96% improvement in 𝑚𝑅@1, and an 8.95% reduction in 𝐹.
LifelongPR, rely on replay sample selection considerng information quantity and spatial distribution, and continual learning framework based on prompt learning, achieves better performance for Continual Learning for Point Cloud Place Recognition. (a) Overall workflow, (b) Replay sample selection, (c) Prompt module. 1)The replay sample selection method dynamically allocates sample sizes according to each dataset’s information quantity and selects spatially diverse samples for maximal representativeness. 2)The prompt learning-based continual learning framework consists of a lightweight prompt module and a two-stage training strategy, enabling domain-specific feature adaptation while minimizing forgetting.
