基于Alphapose骨骼点和GRU的摔倒检测技术

Fall Detection Technology Based on Alphapose Bone Key Points and GRU Neural Network

摘要: 基于现有的一些人体摔倒检测模型实现复杂，适用性差等缺点，本文提出了一种新的更加简便，适用性更强的人体摔倒检测模型。该模型是一种基于GRU神经网络和人体骨骼关键点的人体摔倒检测模型。该模型中，先通过Alphapose对每一帧图像进行人体骨骼关键点识别与检测，然后将得到的骨骼关键点数据进行归一化数据处理，再分组输入到GRU神经网络中进行时序特征提取，最后将GRU模型中隐含层的输出向量输入到全连接层进行处理并得出检测结果。本文使用的是UR Fall Detection Dataset热舒夫大学摔倒数据集进行测试实验，并与多种检测模型的实验性能进行了横向对比。实验结果表明本文的模型在多场景，多视角和多种摔倒姿势等情况下较其他模型都有较高的检测精度，且实现难度较其他模型而言要低。

Abstract: Based on the shortcomings of some existing human fall detection models, such as complex implementation and poor applicability, this paper proposes a new simpler and more applicable human fall detection model. This model is a human fall detection model based on GRU neural network and key points of human bones. In this model, the key points of human bones are identified and detected through Alphapose for each frame of image, Then the obtained bone key point data is normalized data processing, and then grouped and input into the GRU neural network for time series feature extraction, finally, the output vector of the hidden layer in the GRU model is input to the fully connected layer for processing and the detection result is obtained. This article uses the UR Fall Detection Dataset Rzeszow University fall data set for test experiments, and horizontal comparison with the experimental performance of a variety of detection models. Experimental results show that the model in this paper has higher detection accuracy than other models in multiple scenes, multiple perspectives, and multiple falling postures, and the difficulty of implementation is lower than other models.