Localized Trajectories for 2D and 3D Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human Action Recognition

Background:

• Dense Trajectories (DT) is a successful action recognition method, but susceptible to noise and background motion.
• Irrelevant trajectories can degrade the performance of action recognition systems.

Purpose of the Study:

• To propose Localized Trajectories (LT) as an improvement over Dense Trajectories for action recognition.
• To enhance action representation by focusing on human body joint movements.
• To generalize LT to 3D using depth information.

Main Methods:

• Clustering motion trajectories around human body joints detected by RGB-D cameras.
• Encoding clustered trajectories using a local Bag-of-Words model.
• Generalizing LT to 3D (3D LT) by incorporating depth data for radial displacement analysis.

Main Results:

• LT provides a more discriminative representation of human actions compared to standard DT.
• 3D LT captures motion perpendicular to the image plane, offering richer spatial-temporal information.
• Extensive experiments on five datasets demonstrate the effectiveness of the proposed methods.

Conclusions:

• Localized Trajectories offer a robust and accurate approach to action recognition, particularly in cluttered environments.
• The 3D extension of LT further improves performance by leveraging depth information.
• The proposed method advances the state-of-the-art in human activity analysis.