Broad Feature Alignment for Robotic Ground Classification in Dynamic Environment

摘要

Due to the potential hazards of traversing the field ground, the robotic ground classification (RGC) has been widely concerned and extensively studied in environmental perception tasks. However, the RGC usually performs well in experimental environment (source domain) but poorly in working environment (target domain), as the environmental change might cause the ground properties variation and consequently the data drift that comes down to the domain adaptation problem. Hence, we propose the broad feature alignment to suppress the deterioration in accuracy of RGC upon dynamic environment. The contribution is threefold, first, the features are represented first via a broad learning network to improve the feature alignment performance; second, the target domain information preserving term is adopted to design a specific alignment object for suppressing data drift, so that the source-domain broad features could be aligned to the designed alignment object via projected maximum mean discrepancy; third, the feature-temporal manifold regularizer is exploited to improve the alignment consistency of source-domain represented features. The proposed method is verified experimentally on the data gathered by a microtracked robot.