End-to-End Active Object Tracking and Its Real-World Deployment via Reinforcement Learning

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Conventional active object tracking separates tracking and control, hindering joint tuning and requiring extensive manual effort.
• Existing methods are often inefficient due to complex tuning and real-world trial-and-error processes.

Purpose of the Study:

• To develop an end-to-end solution for active object tracking using deep reinforcement learning.
• To overcome limitations of conventional methods by enabling direct frame-to-action prediction.

Main Methods:

• An end-to-end deep reinforcement learning framework utilizing a ConvNet-LSTM function approximator for direct frame-to-action prediction.
• Implementation of environment augmentation techniques and a customized reward function for effective training.
• Training conducted in simulators (ViZDoom, Unreal Engine) with subsequent evaluation on real-world data.

Main Results:

• The proposed tracker shows strong generalization across unseen object paths, appearances, backgrounds, and distracting objects.
• The system demonstrates robustness, including the ability to recover from occasional target loss.
• Successful transfer of tracking capabilities from simulation to real-world scenarios, validated on the VOT dataset and a physical robot.

Conclusions:

• Deep reinforcement learning offers an effective end-to-end solution for active object tracking.
• The proposed training techniques and environment augmentation are critical for successful tracker development.
• Simulation-based training shows significant potential for real-world robotic applications, reducing the need for extensive real-world data collection and tuning.