Research On Indoor Fall Detection Technology Based On Wireless Signal

Recently,aging population has intensified,and falls have been one major threat to the health of elders.Thus,how to automatically detect falls and treat the fall human timely become critical to protect the health of elders.Wearable sensors-based,ambient devicebased and computer vision-based solutions have been proposed to implement highprecision fall detection with some limitations.Therefore,wireless sensing technology has drawn a lot of attention due to its non-contact and non-line of sight.However,the multipath effects,which commonly exist in indoor environments,have serious impact on the receiver performance,leading to the difficulty of extracting falls-related feature,thus bringing a great challenge to human fall detection.In this thesis,we propose an indoor fall detection framework based on wireless signals to detect falls and make innovations from four aspects: network structure,data augmentation,model migration and real-time system.(1)We first collect the channel state information(CSI)from the commodity Wi Fi device when the person is performing activities.Then,the CSI information is preprocessed to denoise.Furthermore,a deep learning model based on long-term and short-term memory network is trained to extract fall related features from wireless signals to detect falls.(2)Due to the multi-path transmission mode of wireless signals,it is difficult to extract features that there is a complex mapping relationship between human activities and wireless signals.Therefore,we propose to use multi-scale method and multi-layer classifier method for feature extraction,which optimizes the network structure and improves the detection performance of the network.(3)Because the data-driven deep learning method needs a large amount of data,and the cost of wireless data acquisition and annotation is too high.We propose a data augmentation method of wireless signal to improves the overall detection performance.At the same time,semi-supervised learning is applied to the task to reduce the pressure of building datasets.(4)It is difficult to transfer the trained model to the new environment for the wireless signal is sensitive to environmental changes.Therefor,we propose the pseudo label method,which can detect falls in the new environment without the data with labels in the new environment.(5)When applied to real scenes,the requirements of continuous,free activity and real-time pose challenges to the system.By constructing the real-time processing system and optimizing the data extraction method,the real-time and continuity of the system are ensured.