Trajectory Predictor by Using Recurrent Neural Networks in Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Motion models are critical for visual tracking, with particle filters and sliding windows being common.
• Existing methods often lack the ability to transfer learned motion knowledge between objects.

Purpose of the Study:

• To propose a novel trajectory predictor for object motion estimation using deep learning.
• To enable knowledge transfer from annotated trajectories to new objects for improved tracking.
• To develop a robust method for scenarios with limited annotated data.

Main Methods:

• Utilizing convolutional neural networks (CNNs) for visual feature extraction of target objects.
• Employing a long short-term memory (LSTM) model to predict motion based on visual features and trajectory data.
• Introducing a dynamic weighted motion model combining the trajectory predictor with a random sampler for data-scarce scenarios.

Main Results:

• The proposed trajectory predictor effectively learns and transfers motion priors.
• Experiments on a real-world vehicle dataset and a benchmark dataset demonstrate superior performance.
• The method outperforms several state-of-the-art trackers in motion prediction accuracy.

Conclusions:

• Deep learning-based trajectory prediction offers significant advantages for visual tracking.
• The proposed method enhances tracking accuracy and robustness, especially in challenging conditions.
• Knowledge transfer in motion modeling is feasible and beneficial for object tracking applications.