Pet Dog Remote Monitoring System Based On Machine Vision

With the growth of per capita GDP,people's living and consumption levels have greatly improved.More and more families choose to keep pets.The role of pets in the family has also changed from “companion” to “children”,and the consumption model has approached the mother and infant market.Then,most office workers will face the distress of being unable to monitor and accompany their pets.Therefore,the design of remote pet monitoring system in this paper has certain practical value.The thesis mainly designs an embedded remote monitoring system for pet dogs.The system hardware includes the server processor Jetson TX2,an external USB camera,a USB amplifier,and an independently designed client APP.The embedded server mainly completes the video stream acquisition and preprocesses it,then uses target detection,behavior recognition and classification technology in deep learning to identify the behavior of pet dogs,sends the pet dog status information to the client APP,and gives a real-time vibration reminder when abnormal mania and other states are recognized,so as to the owner can browse the pet dog video and confirm the current status of the pet dog in real time,and make voice reminders to their dog through APP,which are useful to solve the owner's problems of real-time understanding pet status,interacting with the pet,and implementing remote care of the pet stayed at home.Specific research work is:1.An image zoom comparison experiment using the Python language to implement the bilinear interpolation algorithm and the bicubic Ferguson surface interpolation algorithm,verifying the excellent scaling effect of the bicubic Ferguson surface interpolation algorithm.It is beneficial to obtain details of pet behaviors and has been applied in the preprocessing of pet dog video frame images;2.Considering that it is convenient for the real-time video stream processing on the embedded server-side,in the pet dog detection neural network,the model optimization technology of channel pruning and layer pruning is studied,and theYOLOV3 network implemented by Pytorch is lightly trained to compress the model parameters by 98% and the model is 3.6MB after pruning.And under the premise of ensuring detection accuracy,the inference time is reduced to a half of the original network;3.The lightweight network Shuffle Net V2 is used to recognize five types of behaviors of the dog,such as sitting,standing,lying,mania,and eating.Combined with above target detection network to achieve a “pet dog detection + behavior recognition” network,with an accuracy rate of 89.5%.The recognition result is sent to the client APP through the Socket mechanism;4.According to the system requirement,the APP Android client module is designed and implemented,through which one can complete the functions of the real-time video browsing,voice playback control,real-time storage of pet dog behavior and interact with pet dog;5.The pet dog remote monitoring system on the Jetson TX2 embedded platform is realized,The embedded hardware TX2 is used as the server to running “pet dog detection + behavior recognition” network.Among them,the real-time video stream is collected by USB camera connected to TX2,the transmission of the video stream is realized by the video streaming server MJPG-streamer.And the peanut shell mapping method is used to enable real-time data communication between the external APP and the server connected to the internal network,thereby completing the actual test verification.Experimental test results show that the system has completed the expected function of pet dog monitoring,and the proposed lightweight behavior recognition model can achieve a processing frame rate of 25 fps while guaranteeing 89.5%accuracy of behavior recognition on the TX2 side,realizing the remote monitoring and real-time behavior recognition of pet dogs.