基于AlphaPose与改进LightGBM算法的触电跌倒检测方法

Electric Shock Fall Detection Method Based on AlphaPose and Improved LightGBM Algorithm

针对电力应用场景下人员触电跌倒的问题,文章提出了一种基于AlphaPose与自适应特征权重LightGBM算法的触电跌倒检测方法。该方法首先利用AlphaPose对人体骨骼关键点进行检测,接着根据人体骨骼关键点数据和人体检测框数据计算获得8种人体姿态时空特征。进一步对传统LightGBM算法进行改进,根据特征重要性为特征向量加权,然后训练得到自适应特征权重LightGBM分类器,并将8种人体姿态时空特征输入自适应特征权重LightGBM分类器判断是否为跌倒行为。对Le2i公开跌倒数据集和自制数据集进行实验,并与多种检测模型对比,实验结果表明该方法在复杂度、判断准确性等方面较传统检测模型有显著提升。

Aiming at the problem of people's electric shock and fall in power application scenarios,this paper proposes an electric shock fall detection method based on AlphaPose and adaptive feature weight LightGBM algorithm.Firstly,the bone key points are detected by AlphaPose.Then,eight space-time features of human posture are calculated by the data of skeleton keypoints and human body detection frame data.Next,traditional LightGBM algorithm is improved,and the feature vector is weighted according to the importance of feature,and the adaptive feature weight LightGBM classifier is obtained by training.Finally,eight space-time features of human posture are input into the adaptive feature weight LightGBM algorithm to detect the falling behaviors.Through comparisons with other detection models on Le2i public fall dataset and selfmade dataset,the experimental results show that this method has a significant improvement over the traditional detection model in terms of complexity and judgment accuracy.