Achieving Real-time Visual Tracking with Low-Cost Edge AI
Van Minh Do, Meiqing Wu, Siew-Kei Lam, Thambipillai Srikanthan · 2024
Visual multiple object tracking (MOT) algorithms based on deep learning are computationally intensive, and often cannot achieve real-time performance on low-cost edge computing platforms. We propose an algorithmic-hardware co-design methodology that combines novel algorithm augmentations and architecture mapping of state-of-the-art visual MOT on heterogeneous multi-core processor. We applied the proposed algorithm augmentations to two deep visual MOT pipelines. Experiments based on widely-used datasets demonstrate that the proposed methods outperform the baselines. We also show that the proposed methodology is able to achieve high performance on a low-cost embedded device (Odroid N2+), making it viable for real-time automated traffic surveillance with edge AI.