Research On Infant Kicking Quilt State Detection Technology Based On YOLOv4

Infants and young children are easy to kick the quilt and fail to find it in time during sleep,which is easy to cause colds and affect their parents’ sleep quality to a great extent.Therefore,it is very important to detect the kicking state of high-quality infants in real time.At present,the common infant kicking quilt state detection technology is based on sensors,such as temperature sensor,infrared sensor,photoelectric sensor,etc.Although the implementation is simple,the detection accuracy is easily affected by ambient temperature and light.The detector needs to be fixed on the body,so it is very inconvenient and has potential safety hazards.In addition,due to the lag of sensor response,the real-time performance of detection is poor.The kick quilt state detection based on vision has the advantages of non-contact and visible to human eyes.With the development of detection methods,it has gradually become the mainstream.At present,the traditional visual detection is easily affected by light,infant clothing,quilt patterns and so on,and its algorithm has poor robustness.In recent years,the rapid development of deep learning method can effectively avoid the interference of the external environment and has high recognition accuracy,which makes it a hot spot in the field of target detection research.However,due to its complex network model and the need for a large number of samples for training,it is difficult to deploy to the embedded platform.Therefore,this paper studies such problems,and the main work is as follows.At present,the sample labeling rules in deep learning adopt unstructured constraints,which is easy to cause the loss of structural information in the image target,thus affecting the detection accuracy.Therefore,this paper proposes an image annotation rule based on human structure constraints,which improves the detection accuracy of key parts by adding human structure constraints to key parts.In addition,considering that there are too few samples in the actual application scene of infant kicking state detection,it is difficult to train the model of Yolo V4 algorithm.Therefore,this paper proposes a small sample training method based on incremental learning,which is realized by introducing a super parameter to adjust the proportion of source training set and target training set in batch training.Experiments show that the proposed method can better adapt to the application scenario of this paper.In addition,considering the large amount of parameters and calculation of Yolo V4 model,it is difficult to deploy to the embedded platform with limited storage space and computing power.This paper presents a lightweight detection method of infant kicking quilt state based on Yolo V4 model.The model lightweight in the algorithm is realized by controlling the depth and width of the network,introducing small cross residual modules and removing redundant detection heads,which can be conducive to deployment on the embedded platform.Experiments show that this method maintains the same accuracy as Yolo V4,and the model weight is 1.8 times smaller than Yolo V4 tiny.Finally,in order to verify the feasibility of the proposed algorithm,a prototype system for infant kicking quilt state detection is designed and implemented.The prototype system can realize the functions of user-defined kicking state,real-time kicking state detection,kicking alarm,saving detection results and so on.