Abstract:
This paper addresses the challenge of single-object tracking on resource-constrained devices, a critical aspect for applications like autonomous drones and robotics. We propose an efficient real-time tracking system that leverages the strengths of transformer-based neural networks in combination with correlation filters. Our research makes several key contributions: first, we conduct a comprehensive analysis of existing object tracking algorithms, identifying their advantages and limitations in resource-constrained environments. Second, we develop a novel hybrid tracking system that seamlessly integrates both neural networks and traditional correlation filters. This hybrid system is designed with a switching mechanism based on perceptual hashing, which allows it to alternate between fast but less accurate correlation filters and slower but more accurate neural network-based algorithms. To validate our approach, we implement and test the system on the Jetson Orin platform, which is representative of edge computing devices commonly used in real-world applications. Our experimental results demonstrate that the proposed system can achieve significant improvements in tracking speed while maintaining high accuracy, thereby making it a viable solution for real-time object tracking on devices with limited computational resources. This work paves the way for more advanced and efficient tracking systems in environments where computational power and energy are at a premium.
Fulltext:
PDF file (961 kB)