Elderly Fall Detection Based On Improved SSD Algorithm

With the increasing empty nest phenomenon,the health problems of the elderly have attracted much attention.Falling is one of the main reasons affecting the health of the elderly.If the elderly living alone do not get help and treatment for a long time after falling,it will cause serious consequences.Therefore,the research on the detection of falls in the elderly is of great significance.This paper mainly studies the fall detection method based on vision.Because the Single Shot Multibox Detector（SSD algorithm）in target detection uses multi-scale feature map to predict and has a good detection effect on small targets,a lightweight network model is proposed based on SSD algorithm.The model can directly get the detection results by inputting the falling video,so as to prevent the loss of information in the detection process.Moreover,the structure of the model is simple and the amount of calculation is small,which speeds up the detection speed.This paper conducts experiments on the public dataset of multiple cameras fall dataset after annotation processing.The results show that our model has better detection effect,and can be applied to monitoring equipment to facilitate the fall detection of the elderly living alone.The main work of this paper is as follows:1)In the preprocessing part,this paper uses image random clipping and multi-scale training to increase the representation of samples and improve the performance of the model.At the same time,we use the generated random size pictures to alleviate the lack of computational power.2)In order to improve the detection accuracy and speed up the detection speed,an end-to-end lightweight network model is proposed in this paper.Firstly,VGG16 backbone network in SSD network is modified to mobilenetv2 network.Then,considering that the detection is carried out in the indoor scene and the size of the detection target is not particularly small,the up sampling part and the prediction output branch corresponding to the third down sampling are removed.Finally,in the last two stages of the network,the FPN structure is introduced for up sampling,which integrates the information of low-level and high-level features to enrich the information.The up sampling results are processed with the context module in the SSH algorithm（single stage headless face detector）,which increases the effective receptive field and makes the detection of small targets more accurate.3)This paper optimizes the loss function in SSD algorithm.Because the CIo U loss function has scale invariance,the detection effect of the detector is more balanced and can accurately detect the position of small objects.Therefore,we replace the smoothl1 loss function of location loss in SSD algorithm with CIoU loss function.