基于PointNextMamba联合网络的毫米波点云人体行为识别方法

A millimeter-wave point cloud human behavior recognition method based on PointNextMamba joint network

摘要:

针对现有的毫米波雷达人体行为识别方法依赖关键点提取或固定帧融合策略导致性能受限、鲁棒性不足及计算复杂度高的问题，提出了一种基于PointNextMamba联合网络的人体行为识别方法。该方法对毫米波雷达原始信号进行处理，生成并动态融合多帧点云，利用PointNextMamba网络提取时空特征并完成分类。与两阶段方法相比，避免了中间表示依赖；与其他一阶段方法相比，提升了动态场景鲁棒性与细节特征捕捉能力，同时降低了计算复杂度，与UWB-PointTransformer的o(n2)相比，本方法时间复杂度为o(n)，在GPU上的平均推理速度提升了5倍。在Vayyar数据集上的消融实验及实际场景测试表明，该方法在识别准确率与效率方面均优于现有方法，能够有效满足实际应用中对实时性与鲁棒性的需求。

Abstract:

To address the performance limitations, robustness limitations, and high computational complexity of existing millimeter-wave radar human action recognition methods, which rely on keypoint extraction or fixed-frame fusion strategies, a human action recognition method based on a PointNextMamba joint network is proposed. This method processes the raw millimeter-wave radar signal, generates and dynamically fuses multi-frame point clouds, and utilizes the PointNextMamba network to extract spatiotemporal features and perform classification. Compared with two-stage methods, this method avoids reliance on intermediate representations. Compared with other one-stage methods, it improves robustness in dynamic scenes and the ability to capture detailed features while reducing computational complexity. Compared with the o(n2) time complexity of the UWB-PointTransformer, this method has a time complexity of o(n), resulting in a fivefold increase in average inference speed on a GPU. Ablation experiments on the Vayyar dataset and real-world scenario testing demonstrate that this method outperforms existing methods in both recognition accuracy and efficiency, effectively meeting the real-time and robustness requirements of practical applications.