A Performance of Low-Cost NVIDIA Jetson Nano Embedded System in the Real-Time Siamese Single Object Tracking: A Comparison Study

The tracking of objects is a complex mission computer vision and machine learning (ML). There are several types of objects tracking like tracking one object or more than one. The tracking of one object is applied in video frames as the tracking of more than one object is applied to tracking for multiple objects in the video. Single object tracking is usually implemented using the method of correlation filter-based or of Siamese Network-based. Siamese Network, the state of art method, has an active search area, nowadays, due to its good achievements in accordance with localization real-time and accuracy application. Especially within a new surveillance system that is built on UAV to get unbounded tracking. GPU-based embedded systems give superior performance in comparison with CPU-based systems in the implementation of ML in the terms of speed. In this paper low-cost NVIDIA Jetson nano embedded system performance was evaluated for real-time Siamese single object tracking. 14 Siamese single object tracking algorithms were tested using NVIDIA Jetson nano board. The result shows that the bord gives the best performance with the Lighttrack algorithm with 8.3 frames per second speed. Such performance can be used in real-time tracking applications.