HNST: Hybrid Neural State Tracker for High Speed Tracking
Zhe Zou, Yujie Wu, Rong Zhao · 2021
Object tracking is a fundamental problem in perception, which is of great importance for intelligent robots. Recent years have witnessed breakthroughs of deep neural networks (DNN) in intelligent perception with a remarkable improvement on tracking accuracy. However, the huge computing cost and high latency restrict their applications in embedded robotic platform. In contrast, traditional correlation filters-based trackers use handcraft low-level features with higher tracking speed and lighter computational overhead. Both methods have their own advantages. Integrating them may provide a complementary strategy to achieve flexibility between speed and accuracy, which promises to handle more complex scenarios. This work proposes a hybrid neural state tracker (HNST) that combines DNN-based detection and Kernelized Correlation Filter (KCF) tracking using a policy to determine when to activate detection and whether to update the template of tracking target. Furthermore, we develop a similarity discrimination and establish a hybrid neural state machine for decision-making. Through comprehensive evaluation of the speed and accuracy of multiple standard video datasets, our approach improves more than 70% on precision and 95% on AUC with real-time tracking speed (~100 FPS) on CPU over the KCF tracker baseline, indicating the great potential of HNST in high-speed tracking.