| With the continuous development of deep learning technology,more and more researchers have begun to focus on how to apply deep learning models to practical scenarios,especially in resource-constrained embedded devices.In these areas,efficient deployment and real-time performance of models become crucial.This article aims to study lightweight deep learning models,including the design of lightweight target detection and tracking networks and mobile deployment,to provide feasible solutions for practical applications.The main work of this article includes the following three aspects:Firstly,based on the commonly used object detection network YOLOv5,an improved network based on Ghost convolution is proposed.The Ghost convolution structure is used to replace the original convolution structure,which can significantly reduce the number of parameters and computational complexity of the model,thereby improving the inference speed of the model while maintaining high detection accuracy.During the training process of the network,the SIoU loss function is used for optimization.Compared with the traditional IoU loss function,the SIoU loss function can better handle problems such as incomplete occlusion,thereby improving the accuracy of the network.Finally,the model’s parameter quantity is reduced by 47% and the inference speed on embedded platforms is improved by 20%,while only reducing the model’s accuracy by about 5%.Secondly,a complete embedded object detection application is designed using Ascend 310 as the neural network computing unit.The project requirements are analyzed and the application’s various modules are designed.The communication process with the upper computer and the model inference are decoupled through modular design.The functions of each module are implemented separately.The performance of the application is optimized based on the dedicated data processing unit DVPP and quantized models,ultimately achieving an inference speed of over 20 FPS.Finally,based on Ascend 310 artificial intelligence acceleration chip,a target tracking application combined with object detection is designed,and the improved GhostYOLOv5 and SiamRPN++ algorithms are implemented on embedded platforms and optimized for performance.By using performance tests on relevant target tracking datasets,a real-time embedded visual application that meets the requirements for tracking is achieved. |
