Real-Time Object Detection And Embedded Implementation Based On Lightweight Network

The proposed management model of "smart city" means that object detection based on deep learning will play an important role in urban development.However,most of the current object detection models are large in size and requires a lot of computing resources,which is difficult to deploy to mobile devices.Therefore,this paper achieves a balance between speed and accuracy by making lightweight improvements to different target detection models,on the premise that the size of the model meets the requirements of the embedded platform.And deploy the target detection model based on deep learning on different embedded platforms to realize the application of object detection technology on the embedded development boards.The main work and innovations of this paper are as follows:(1)A lightweight face detection network based on depthwise separable convolution and attention model is proposed.With SSD as the baseline model,a balance of accuracy and speed is achieved.In order to improve the accuracy of the model,Receptive Fields Block,attention model,combination of DIo U and Non-Maximum Suppression,and Mish activation function are introduced into the algorithm;In order to reduce the size of the model,set the appropriate prior boxes for the scale of the face features,and use depth-separable convolution to replace standard convolutions.Meanwhile,a new dataset was constructed for the complex variation of illumination.The experimental results show that using our dataset can improve the robustness of the model.(2)The YOLOv5 s algorithm based on interactive fusion of multi-scale features is proposed.Using YOLOv5 s as the baseline model,we introduce the Ghost module in the Bottleneck module of backbone to reduce the computational power of the network and reduce the size and computational power of the model.Meanwhile,the original feature collection and redistribution module is introduced and improved.The redistribution of features at different scales enriches the features used for detection and improves the detection accuracy of the model.(3)The object detection algorithm is applied on different embedded platforms.According to the different requirements of the embedded platform,different target detection algorithms are deployed.The deployment process includes steps such as model conversion,environment construction,and model testing.It also improves the object detection algorithm according to the specific acceleration method of different embedded platforms to make it faster on embedded devices.The experimental results show that the two lightweight object detection algorithms proposed in this paper have better results than existing algorithms,and can achieve realtime detection on embedded development boards.