基于骨骼关键点与CNN-LSTM混合网络的溺水行为检测方法研究

A Study on a Drowning Behavior Detection Method Based on Skeletal Landmarks and a Hybrid CNN-LSTM Network

摘要: 溺水是全球范围内导致意外死亡的主要原因之一，尤其对儿童与青少年群体构成严重的公共卫生安全威胁。传统泳池视频监控高度依赖人工值守，普遍存在监视疲劳、漏检率高与应急响应滞后等问题。现有基于纯视觉的溺水检测方法易受水面光照变化、水体反光、水花遮挡等复杂场景干扰，鲁棒性与稳定性不足。针对上述问题，本文提出一种基于骨骼关键点与CNN-LSTM混合网络的溺水行为检测方法。该方法首先采用姿态估计模型提取人体骨骼关键点序列，剥离复杂背景与水面干扰，保留纯粹的人体运动结构信息；通过一维卷积网络完成局部姿态空间特征编码，并结合注意力机制强化有效关节特征权重；再利用LSTM网络挖掘姿态时序演变规律，实现溺水异常行为的精准识别。针对真实溺水样本稀缺、数据集难以采集的问题，本文引入合成数据增强策略扩充样本，有效提升模型泛化能力。对比实验与消融实验结果表明，相较于单一CNN模型与LSTM模型，本文所提时空混合网络能够有效捕捉溺水过程中的细微姿态变化，在复杂水域干扰环境下检测精度与鲁棒性显著提升，可有效区分正常游泳与溺水状态。该方法可为智能水域安全监控与自动预警系统提供可靠的技术支撑。

Abstract: Drowning is one of the leading causes of accidental death worldwide and poses a serious public health threat, particularly to children and adolescents. Traditional video surveillance in swimming pools relies heavily on manual monitoring, which is often plagued by issues such as surveillance fatigue, high rates of missed incidents, and delayed emergency responses. Existing purely vision-based drowning detection methods are susceptible to interference from complex scenarios—such as changes in water surface lighting, water reflections, and obstruction by splashes—and lack sufficient robustness and stability. To address these issues, this paper proposes a drowning behavior detection method based on skeletal keypoints and a hybrid CNN-LSTM network. The method first employs a pose estimation model to extract sequences of human skeletal keypoints, filtering out complex background and water surface interference to retain pure human motion structural information; it then uses a one-dimensional convolutional network to encode local pose-space features, combined with an attention mechanism to enhance the weights of effective joint features; an LSTM network is then utilized to uncover the temporal evolution patterns of posture, enabling accurate identification of abnormal drowning behaviors. To address the scarcity of real drowning samples and the difficulty in collecting datasets, this paper introduces a synthetic data augmentation strategy to expand the sample size, effectively enhancing the model’s generalization ability. The experimental results demonstrate that, compared to the single CNN model and LSTM model, the spatiotemporal hybrid network proposed in this paper can effectively capture subtle posture changes during drowning, significantly enhancing detection accuracy and robustness in complex aquatic interference environments. It can effectively distinguish between normal swimming and drowning states. This method can provide reliable technical support for intelligent aquatic safety monitoring and automatic warning systems.