摔倒检测大都依靠传感器设备,此类方法受设备自身和环境因素影响较大,常常无法发挥该有的作用,同时,基于视觉的方法往往实时性较差,鲁棒性不强。针对上述问题,本文提出了一种鲁棒性强、能有效部署在嵌入式设备上的轻量化摔倒检测算法。以YOLOv5为基准模型,首先,融合轻量级注意力机制模块,使网络更关注要识别的目标区域,增强网络的识别精度。其次,使用模型压缩方法对模型进行剪枝,减小模型体积和计算量,使模型轻量化,以提高推理速度和便于部署在嵌入式设备中。最后,对剪枝后的模型进行知识蒸馏,在不提升模型复杂度的前提下提升模型的检测精度。实验结果表明:本文模型相较于基准模型,mAP增加了1.7%,召回率提高了1.2%,模型体积减小了79.1%,浮点运算量降低了70.9%。将本文模型部署在嵌入式设备Jetson Nano上,检测速率达到13.2 frame/s,基本满足实时性摔倒检测的要求。

Fall detection mostly depends on sensor equipment.The method is highly influenced by equipment and environmental factors,and often can not work well.In addition,vision-based methods are often not effective in terms of real-time and robust. In order to solve these problems,a lightweight fall detection algorithm is proposed with strong robustness and convenient deployment in embedded devices.Taking YOLOv5 as the benchmark model,the lightweight attention mechanism module is firstly integrated to make the network focus on the target area to be identified and enhance the recognition accuracy of the network.Secondly,the model is pruned by the model compression method,which reduces the volume and calculation.Therefore it makes the model lightweight,so as to improve the reasoning speed and facilitate deployment in embedded devices.Finally,knowledge distillation is carried out on the pruned model,which can improve the detection accuracy without increasing the complexity of the model.The experimental results show that compared with the benchmark model, the mAP of this model is increased by 1.7%,the recall is increased by 1.2%,the model volume is reduced by 79.1%,and the floating-point operation is reduced by 70.9%.The proposed model is deployed on the embedded device Jetson Nano,and the detection speed is up to 13.2 frame/s,which basically meets the requirements of real-time fall detection.

国家自然科学基金(61901059,51978079)资助项目