针对多传感器人体行为识别中存在的异构数据融合难题与实时性不足等问题,提出了一种融合双向长短期记忆网络（BiLSTM）与门控注意力机制的多传感器融合模型。该方法通过BiLSTM捕捉双向长时序依赖,利用门控注意力实现多源特征自适应加权,从而提升特征表达能力与融合效率。实验表明,模型在UCI-HAR数据集上的最高准确率达95.3%,同时在参数量与单样本推理延迟方面较轻量化Transformer模型更具优势,在保证识别精度的前提下实现了更优的效率表现。消融实验进一步验证了BiLSTM与注意力机制对性能提升的关键作用。本研究为复杂场景下的行为识别提供了一种兼具识别精度与推理效率的解决方案,具备良好的实际应用潜力。
Abstract
A multi-sensor fusion model based on Bidirectional Long Short-Term Memory (BiLSTM) and gated attention mechanism is proposed to address the challenges of heterogeneous data fusion and insufficient real-time performance in multi-sensor human activity recognition. This method captures bidirectional long-term dependencies through BiLSTM and employs gated attention to achieve adaptive weighting of multi-source features, effectively enhancing feature representation capability and fusion efficiency. Experimental results demonstrate that the proposed approach achieves a highest accuracy of 95.3%,in addition, under a comparable lightweight setting, the proposed model exhibits lower parameter count and faster inference latency than the Transformer baseline, thereby achieving a more favorable balance between accuracy and efficiency.Ablation studies further confirm the critical role of BiLSTM and the attention mechanism in improving performance. This research provides a solution that balances recognition accuracy and inference efficiency for behavior recognition in complex scenarios, showing strong potential for practical applications.
关键词
多传感器融合 /
注意力机制 /
BiLSTM /
人体行为识别
Key words
multi-sensor fusion /
attention mechanism /
BiLSTM /
human activity recognition (HAR)
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] 余金锁,卢先领.基于分割注意力的特征融合CNN-BiLSTM人体行为识别算法[J].电子测量与仪器学报,2022,36(2): 89-95.
[2] 武东辉,许静,陈继斌,等.基于融合注意力机制与CNN-LSTM的人体行为识别算法[J].科学技术与工程,2023,23(2): 681-689.
[3] 宦若虹,葛罗棋,吴炜.基于多特征融合CNN-BLSTM的人体复杂行为识别方法: CN111125984A[P].2020-04-29.
[4] Lee D,Min S D.Wearable Sensor-Based Human Activity Recognition with Transformer Model[J].Sensors,2022,22(18):6913.
[5] Djenouri Y,Fehmida A,Gatti S,et al.Hybrid and Lightweight Transformer for Efficient Human Activity Recognition[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops.2023:1-9.
[6] Karimi N,Baghshah M S.Transformer-based Deep Reverse Attention for Human Action Analysis[J].Expert Systems with Applications,2023,225:120086.
[7] Shuvo S B,Siddiqui J,Hasan M R,et al.Transformer-based deep reverse attention for multi-sensory human activity recognition[J].Information Fusion,2023,92:424-439.
[8] Lee S,Cho H.Discrete Representation Learning for Wearable HAR[J].Sensors,2024,24(3):1157.
[9] Yuan L,Ma Y,Bach S,et al.Large-Scale Self-Supervised Learning for Wearable-Sensor HAR[J].npj Digital Medicine,2022,5:110.
[10] Reyes-Ortiz J-L,Oneto L,Samà A.Multimodal Sensor Fusion for HAR:A Review[J].Information Fusion,2022,80:241-259.
[11] Sun Y,Han B Q,Wang S T,Liu Y.Progressive neural architecture search for mobile human activity recognition[J].计算机应用,2022,42(12):3534-3544.
基金
广东省教育厅特色创新项目（自然科学）“多传感器融合与逻辑判断算法在居家养老摔倒检测与告警系统中的研究与应用”,项目编号：2023KTSCX190; 广东白云学院重点科研项目,项目编号：2023BYKYZ02
