基于国产嵌入式平台的实时目标跟踪算法OA

To address the challenge of deploying existing visual object tracking algorithms on embedded platforms with limited hardware resources,this paper proposes a real-time object tracking algorithm,SiamRT,tailored for the domestic embedded platform RK3588.The design incorporates a lightweight Siamese backbone based on VGG16 to reduce network complexity and enhance feature depth,capitalizing on NPU acceleration characteristics.A channel adjustment layer is employed to solidify the cross-correlation parameters while constructing a multi-task feature fusion network.This enables customized feature fusion ratios for classification and regression sub-tasks.Additionally,a lightweight corner prediction sub-network based on the Ghost mechanism is introduced to eliminate background redundancy and improve localization accuracy.A distributed deployment strategy is utilized,designing an integrated detection and tracking model combining YoloV5 and SiamRT.This approach substitutes manual selection of initial frame targets with a detector and optimizes the practical application process of the algorithm through multi-threading.The algorithm achieves an accuracy of 89.7%and a success rate of 69.6%in the OTB100 benchmark test and demonstrates robust performance on several large-scale public datasets.Deployed on the embedded platform,it achieves a running speed of 63 frames per second.These data and simulated results in real scenarios confirm the algorithm's significant military system application value.It holds strategic significance in enhancing the intelligence of national defense weaponry and equipment.