Design And Implementation Of An IoT-Based Lightweight Target Detection Platform

With the gradual advancement of intelligent regulation of public transport,as well as the rapid development of the Internet of Things and the field of lightweight target detection,it is becoming increasingly clear that people’s production and lives are being affected.However,it is difficult for many SMEs and individual researchers to go deeper into practice due to industrial scenario limitations and expertise.In addition,considering the wide distribution location of public transport and the time delay caused by the transmission of collected data to the cloud for processing,this paper designs an IoT-based lightweight target detection platform for target detection application scenarios,empowering part of the detection computation to embedded devices and realising functions such as busassisted driving,passenger flow statistics,data visualisation and intelligent processing.The main work includes:(1)A lightweight target detection platform is designed and implemented.This paper first analyses the overall objective and overall architecture design of the platform,and then divides the whole platform system into dataset management module,model training and conversion module,model evaluation module,model quantification and deployment module,and real-time detection result monitoring module according to the functions.(2)A lightweight target detection network based on YOLO is designed and implemented.To address the problem of model lightweighting,the huge Darknet backbone network in the traditional YOLO algorithm is replaced by MobileNet,and the redundancy of channels is measured by means of Euclidean distance,and channel deletion operations are carried out to reduce the scale of the network and the memory occupation of the model.(3)A lightweight SSD-based target detection network micro-SSDBasicRFB is designed and implemented.the network model uses four branches for prediction,incorporates the core ideas of MobileNet,uses deep separable convolution,and designs the number of convolutional layers and channels according to the needs of practical application scenarios,and optimises the model size and detection speed.Optimisation.Finally,to further increase the detection accuracy,an improved BasicRFB layer is added to maintain a good balance between detection accuracy and speed.(4)An experimental validation platform is built to validate the platform and algorithms designed and implemented in this paper on an embedded platform for application scenarios,and the software design for multi-model fusion detection and communication transmission is designed and implemented on the embedded device.