A lightweight single object tracking algorithm for low-power embedded devices
Open Access
- 2 February 2024
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 642-648
Abstract
In order to achieve higher tracking accuracy and robustness, the parameter and computational complexity of single object tracking algorithms have become increasingly large, which contradicts the practical application scenarios of strictly limiting computing resources and power consumption. To address this issue, we design a lightweight improved algorithm based on SiamFC. Specifically, we add BlurPooling module to the feature extraction network to reduce information loss caused by downsampling. At the same time, we use the Blackman window with a large main lobe width and severe side lobe attenuation as the window function, which allows the tracker to pay more attention to the effective information extended by blur operations and reducing attention to surrounding noise. Finally, we analyzed the parameter quantity and computational complexity and conducted experiments on benchmarks. Our method can improve tracking precision by 1.3% on UAV123 without increasing the parameter quantity.Keywords
This publication has 7 references indexed in Scilit:
- Deep Learning for Visual Tracking: A Comprehensive SurveyIEEE Transactions on Intelligent Transportation Systems, 2021
- Siamese Box Adaptive Network for Visual Tracking2020
- Deformable Siamese Attention Networks for Visual Object Tracking2020
- SiamCAR: Siamese Fully Convolutional Classification and Regression for Visual Tracking2020
- SiamRPN++: Evolution of Siamese Visual Tracking With Very Deep Networks2019
- High-Speed Tracking with Kernelized Correlation FiltersIEEE Transactions on Pattern Analysis and Machine Intelligence, 2014
- The estimation of the gradient of a density function, with applications in pattern recognitionIEEE Transactions on Information Theory, 1975