LightOcean: A Lightweight and Efficient Network for Real-time UAV Tracking.

Haiyang Chen,Weiqiang Wang,Xingzhou Zhang,Wei Zhou,Weisong Shi

DOI: https://doi.org/10.1109/smartworld-uic-atc-scalcom-digitaltwin-pricomp-metaverse56740.2022.00055

2022-01-01

Abstract:Siamese-based trackers have significantly advanced in visual object tracking over the past years. However, most of these trackers emphasize tracking accuracy over efficiency, which limits their real-world deployment on edge platforms with widespread applications such as unmanned aerial vehicles (UAVs). In this paper, we propose LightOcean, a lightweight and efficient aerial tracker. Specifically, we design a dynamic template feature update module and a pixel-level cross-correlation module to improve the robustness and adaptability of the tracker without additional model computation. The former allows the tracker to capture additional space-time information during the tracking process to adapt to the appearance changes of the object. The latter utilizes pixel-level cross-correlation and attention mechanisms to generate similarity maps, improving the accuracy of predicted boundaries. By combining the lightweight and efficient backbone network and the prediction head, LightOcean significantly increases the tracking speed while maintaining accuracy. By running three common UAV benchmarks on Jetson Nano, a typical edge-embedded device, LightOcean presents better performance. The tracking speed is 4x faster than the state-of-the-art tracker, Ocean, while the energy consumption and memory usage are reduced by 80% and 12%. What’s more, the accuracy remained stable, and the precision of LightOcean is 78.6%, which is 2.7% higher than Ocean.