KNN-Based Machine Learning Classifier Used on Deep Learned Spatial Motion Features for Human Action Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human action recognition (HAR) is crucial for surveillance video analysis.
• Existing HAR methods often rely on computationally intensive 3D Convolutional Neural Networks (CNNs) or two-stream networks.
• These heavy networks present challenges in implementation and training due to a large number of parameters.

Purpose of the Study:

• To develop a computationally efficient HAR method.
• To introduce HARNet, a lightweight directed acyclic graph-based residual 2D CNN.
• To create a novel pipeline for generating spatial motion data for action representation learning.

Main Methods:

• Designed HARNet from scratch, a customized lightweight directed acyclic graph-based residual 2D CNN.
• Developed a novel pipeline to construct spatial motion data from raw video input.
• Fed the constructed input into HARNet for simultaneous spatial and motion information processing in a single stream.
• Extracted latent representations for classification using conventional machine learning models.

Main Results:

• HARNet demonstrated superior performance compared to existing state-of-the-art methods.
• Achieved percentage improvements of 2.75% on the UCF101 dataset.
• Showcased improvements of 10.94% on the HMDB51 dataset and 0.18% on the KTH dataset.

Conclusions:

• HARNet offers an effective and efficient solution for human action recognition.
• The proposed spatial motion data pipeline enhances latent representation learning for actions.
• The lightweight design of HARNet alleviates implementation and training challenges associated with heavy deep learning models.