Design And Implementation Of Object Recognition And Alarm System Based On Embedded Platform

Object recognition based on deep learning has broad application prospects in many domains such as safety monitoring,behavioral regulation,and traffic coordination.The current object recognition system still has shortcomings such as magnitude of model,high requirements in hardware,and tremendous cost in deployment,therefore,it is so limited of its application in many domains,especially mobile terminals.With the rapid development of the Internet of Things and edge computing technology,the application requirements of embedded devices are increasing day by day,and the object recognition system based on the embedded platform has become an crucial development trend.However,object recognition systems based on embedded platforms still face great challenges due to low computing ability and small storage space.In this paper,a variety of strategies are used to optimize the traditional object recognition model,and a lightweight human and object recognition network is designed,which greatly reduces the quantity of computation while retaining the accuracy of recognition.An object recognition and alarm system based on embedded platform has been implemented.The main research results and innovations achieved include:(1)A human and object recognition network is constructed,and a model compression method that combines knowledge distillation,sparse training and channel pruning is proposed,which reduces the parameters and disk space of the model by 92.46%and 98.02%respectively;Tested on the collected data set,mAP0.5 is only reduced by 7.6%,and Recall can still reach 79.1%,achieving a balance between detection accuracy and model complexity.(2)Based on the characteristics of embedded devices,a forward inference optimization method that combines linear operation fusion and data accuracy calibration is proposed.The single-frame detection time of the model has been shortened to 0.493 s,which is 49.85%lower than that before acceleration,and 96.37%lower than that the original model.(3)Developed a human and object recognition and alann system based on an embedded platform,and implement the recognition and violation alarm of various targets such as human body,clothing,and firefighting equipment.The core model size of the system is only 1 MB,the average recognition accuracy is over 97%,the response time is within 1 s,and the QPS is 1539.6.