AI Action Recognition Using Multi-Frame Node Sequences
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional action recognition systems rely on dedicated, bulky, and costly cameras, limiting the application range and increasing implementation costs of intelligent interaction scenarios.
Innovation Solution
An AI-based action recognition method that determines action categories for video frames using node sequence information from consecutive frames, employing neural networks to analyze position information and improve recognition accuracy without requiring dedicated cameras.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a dedicated action recognition camera is used, then action recognition accuracy is improved, but device cost and bulkiness increase
Solution Approach 1:
The patent uses a standard camera to capture video data and creates a simplified representation (node sequence information) of the action through computer vision processing. Instead of using an expensive dedicated action recognition camera, the system copies the essential action information from regular video data through skeleton extraction and node sequence generation, achieving accurate action recognition without the bulky dedicated hardware
Solution Approach 2:
The patent replaces the mechanical/dedicated action recognition camera with a standard camera combined with AI-based computer vision processing. The system substitutes physical dedicated hardware with a software-based solution that processes standard video data through neural networks to extract action information, thereby reducing device complexity and cost while maintaining recognition accuracy
2Measurement precision
If dedicated action recognition hardware is used, then recognition accuracy is improved, but implementation cost increases
Solution Approach 1:
The patent employs a standard, inexpensive camera combined with open-source AI algorithms to achieve action recognition. Instead of investing in expensive dedicated hardware, the system uses readily available video data and processes it through neural networks, significantly reducing implementation costs while maintaining acceptable recognition accuracy for intelligent interaction scenarios
3Productivity
If action recognition is performed using single frame data, then processing speed is improved, but recognition accuracy deteriorates
Solution Approach 1:
The patent extracts node sequence information from multiple consecutive video frames in advance, creating a temporal sequence of action representations. By preparing this sequence data beforehand through skeleton extraction and node tracking, the system can efficiently process the action recognition using the pre-extracted multi-frame information, achieving both speed and accuracy
Solution Approach 2:
The patent processes multiple consecutive video frames continuously to extract node sequence information, maintaining temporal continuity of the action data. This continuous extraction of action information across time frames provides the neural network with comprehensive temporal context, improving recognition accuracy while the efficient processing pipeline maintains speed
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
An action identification method and apparatus based on artificial intelligence. The method comprises: determining node sequence information corresponding to video frames in video data to be identified (S201); determining action types respectively corresponding to the video frames according to the node sequence information respectively corresponding to N continuous video frames in said video data (S202); and determining which interaction action an interaction object makes according to the action types (S203). When determining an action type corresponding to a video frame, reference is made to information carried by N continuous video frames where the video frame is, or reference is made to related information of the video frame in past and/or future time, so that more effective information is introduced; even if said video data is obtained by using a non-special action identification acquisition device, relatively high action identification precision can also be achieved in a manner of determining the action types corresponding to the video frames by means of a group of continuous video frames, so that the application range and the implementation cost of an intelligent interaction scenario are reduced.