Lightweight real-time action recognition method based on state decomposition and state machine

By disassembling the action sequence into atomic state and combining deep learning and vector retrieval technology, the shortcomings of existing action recognition methods in terms of recognition accuracy and efficiency are solved, and efficient and accurate action recognition on resource-constrained devices are achieved.

CN119723680AActive Publication Date: 2025-03-28HEFEI ARTIFICIAL INTELLIGENCE & BIG DATA RES INST CO LTD +1
5 Cites 0 Cited by

Patent Information

Application Number
CN202510233515.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-02-28
Publication Date
2025-03-28
Estimated Expiration
2045-02-28

AI Technical Summary

Technical Problem

The existing action recognition methods have significant shortcomings in recognition accuracy and recognition efficiency, especially in resource-constrained devices, and it is difficult to accurately capture the state relationship between targets in complex scenarios.

Method used

The lightweight real-time action recognition method based on state decomposition and state machine is adopted. By disassembling the action sequence into atomic state, combining the dynamic detection of deep learning algorithms and the static detection of vector search algorithms, the recognition accuracy and reliability are improved.

Benefits of technology

It realizes efficient action recognition on resource-constrained devices, improves the accuracy and stability of action recognition in complex scenarios, and can more accurately judge abnormalities and send warning signals in time.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a lightweight real-time motion recognition method based on state decomposition and a state machine, which comprises the following steps: collecting a historical human motion video, performing key frame extraction and preprocessing on the historical human motion video to obtain a historical key frame, and extracting continuous motion segments from the historical key frame according to a time sequence to obtain a motion sequence; disassembling each action in the action sequence into an atomic state to obtain an atomic state information base; performing dynamic detection on the real-time motion video by using the target detection model; constructing a vector retrieval library according to the historical key frames, and performing static detection on the real-time motion video by using a vector retrieval algorithm according to the vector retrieval library; and performing anomaly judgment on the dynamic area detection result and the static area detection result by using a motion state machine, and sending an early warning signal according to an anomaly judgment result. The invention relates to the technical field of action detection, and solves the technical problems of low recognition precision and poor stability of an existing method.
Need to check novelty before this filing date? Find Prior Art