一种基于改进的卷积神经网络人体跌倒检测算法

A Fall Detection Algorithm Based on Improved Convolutional Neural Network

文章针对高质量公开跌倒数据集较少,导致模型泛化能力较弱、检测准确率低、现有网络全连接层参数量过大收敛速度慢的问题,设计了适用于跌倒检测的迁移学习方法,使用GAP(Global Average-Pooling,GAP)层替换全连接层方法,并在隐藏层加入BN(Batch Normalization,BN)操作,优化网络参数,设置了多组对比实验发现改进的网络模型在不同的数据集上训练时间相比于之前有所提升,均取得了不错的效果,使得神经网络既能够在大规模图像数据集上学习通用的特征又能够在公开跌倒数据集上学习跌倒特征,增强了网络的泛化能力。

This article addresses the problems of weak model generalization ability,low detection accuracy,and slow convergence speed due to the limited number of high-quality public fall datasets.A transfer learning method suitable for fall detection is designed,which replaces the fully connected layer method with a Global Average Pooling(GAP)layer and adds a Batch Normalization(BN)operation in the hidden layer to optimize network parameters,Multiple comparative experiments were conducted,and it was found that the improved network model had improved training time on different datasets compared to before,achieving good results.This enabled the neural network to learn both universal features on large-scale image datasets and fall features on publicly available drop datasets,enhancing the network's generalization ability.