Two-Stream 3D CNN Action Recognition for High-Resolution Video

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network systems for video processing face challenges in achieving accurate classification of motion and action recognition, particularly in higher spatial resolution videos, with high computational and memory requirements.

Innovation Solution

A neural network system that combines a first and second 3D spatio-temporal convolutional neural network to process image and optical flow data, utilizing a data combiner to enhance classification accuracy, and employs an inflated 2D neural network architecture for training, with optional pre-training on different datasets to improve performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If 3D spatio-temporal convolutional neural networks are used for video processing, then classification accuracy is improved, but computational complexity and memory requirements increase

Engineering Contradiction:
Improveclassification accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the video processing task into two separate streams: a spatial stream that processes appearance information from video frames, and a temporal stream that processes motion information from optical flow data. Each stream uses its own 3D spatio-temporal convolutional neural network, allowing independent optimization and processing. This segmentation reduces the computational burden on a single network while maintaining high classification accuracy through fusion of both streams.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extends 2D convolutional neural networks to 3D spatio-temporal networks by adding the temporal dimension. This allows the networks to process video data in three dimensions (height, width, time) simultaneously, capturing both spatial features and temporal dynamics. The dimensional extension enables accurate action recognition while the modular two-stream architecture manages the associated computational complexity.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If higher spatial resolution videos are processed, then classification accuracy is improved, but computational requirements increase

Engineering Contradiction:
Improveclassification accuracyVSAvoidcomputational energy
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

By segmenting the processing into spatial and temporal streams, the patent allows each stream to operate at optimized resolutions. The spatial stream can process high-resolution appearance data, while the temporal stream processes motion data that may require less computational energy. This division enables high accuracy processing of high spatial resolution videos while managing energy consumption through specialized processing paths.

Inventive Principle:
Principle #1Segmentation

3Measurement precision

If longer temporal receptive fields are used, then action recognition accuracy is improved, but processing time increases

Engineering Contradiction:
Improveaction recognition accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The temporal stream is specifically designed to handle motion information over time with its own 3D spatio-temporal convolutional network. This dedicated temporal processing path can efficiently capture long-term dependencies and temporal patterns without bottlenecking the overall processing speed. The separation allows the temporal stream to optimize for long receptive fields while the spatial stream handles immediate appearance features, reducing overall processing time.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3602397B1Neural network systems for action recognition in videos
Publication Date: 2026.04.22 GDM HOLDING LLC
  • EP3602397B1 patent drawingFigure 1
  • EP3602397B1 patent drawingFigure 2a
  • EP3602397B1 patent drawingFigure 2b

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing video data. An example system receives video data and generates optical flow data. An image sequence from the video data is provided to a first 3D spatio-temporal convolutional neural network to process the image data in at least three space-time dimensions and to provide a first convolutional neural network output. A corresponding sequence of optical flow image frames is provided to a second 3D spatio- temporal convolutional neural network to process the optical flow data in at least three space- time dimensions and to provide a second convolutional neural network output. The first and second convolutional neural network outputs are combined to provide a system output.