一种融合Alphapose与LSTM的人体跌倒识别模型。

简介：
Numerous Human-Body-Fall-Down Detection models encounter challenges such as reduced adaptability and elevated false detection rates across diverse detection environments. Addressing these limitations, this research introduces a Human-Body-Fall-Down Detection Model leveraging human skeleton keypoints and an LSTM neural network. Specifically, the model utilizes Alphapose to continuously track the skeleton keypoints of the human body across multiple frames. Subsequently, the coordinate sequences derived from these keypoints are segmented into X and Y coordinate sequences, which are then independently fed into an LSTM neural network to capture temporal characteristics. Finally, the output vector from the LSTM’s hidden layer is processed through a fully connected layer to yield the detection results. This study conducted experiments using publicly available datasets – M uHAVi-MAS and Le2i – and compared the performance of this model against a variety of other detection models. The findings demonstrate that this proposed model exhibits notably high detection accuracy in a wide range of scenarios, encompassing various viewpoints and multiple body poses during falls.