An adaptive video semantic analysis method and system based on cognitive feedback closed loop

Abstract

This invention discloses an adaptive video semantic analysis method and system based on cognitive feedback closed loop. The invention first receives a query command and a long-term video stream to be analyzed, initializing a reasoning state memory containing a temporal topology. Next, based on the reasoning state memory, it performs thought chain reasoning on the query command, generating a set of selected time windows, and performs visual sampling and feature encoding on the target time period to obtain a high-dimensional feature vector. Then, it performs fine-grained semantic verification and causal consistency evaluation. Next, it updates the graph topology through the global reasoning state, triggering a closed-loop strategy correction. Finally, when the global confidence level meets a threshold, it generates the final response. This invention reduces computational complexity through hierarchical planning and entropy reduction perception, improves causal understanding and robustness by introducing counterfactual reasoning and temporal logic, achieves privacy protection and bandwidth optimization through feature-level federated interaction, and ensures memory continuity and memory stability by constructing a dynamic event evolution graph and pruning mechanisms.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, computer vision, and multimodal information processing, and in particular to an adaptive video semantic analysis method and system based on cognitive feedback loop, which is applicable to scenarios such as home security with limited computing resources and sensitive privacy, long-term surveillance video retrieval, and intelligent edge computing. Background Technology

[0002] With the widespread adoption of Internet of Things (IoT) technology, the number of video surveillance devices deployed in homes and businesses has grown exponentially, generating massive amounts of unstructured video data. Users often face the daunting task of finding specific details when faced with surveillance footage lasting for hours or even days. For example, a user might need to find "when my pet dog jumped onto the sofa yesterday afternoon" or "the exact moment the delivery person placed the package at the door."

[0003] Existing video semantic analysis technologies mainly suffer from the following bottlenecks:

[0004] The computational resources consumed are enormous: Traditional video analytics paradigms typically employ a "process first, retrieve later" approach, which involves performing frame-by-frame or high-frequency intensive sampling on the video stream and extracting features from all frames using deep neural networks. For long-term surveillance videos, this full-processing method leads to a huge waste of computing power and electricity consumption, making it difficult to deploy on resource-constrained edge devices (such as home security cameras).

[0005] Low retrieval efficiency and lack of temporal logic: Existing keyword-based retrieval methods often can only identify isolated objects or scenes, and struggle to understand natural language queries containing complex temporal logic (such as "first...then..."). Furthermore, indiscriminate full-scale searches result in long retrieval response times, failing to meet the needs of real-time interaction.

[0006] Privacy risks and bandwidth pressure: To achieve high-precision analysis capabilities, existing solutions typically require uploading raw video streams to cloud servers for processing. This not only consumes significant amounts of uplink network bandwidth but also raises serious concerns among users about the potential leakage of their home privacy.

[0007] Therefore, there is an urgent need for a new technological paradigm that can break the shackles of "full-scale perception" and proactively and on-demand allocate perception resources through intelligent cognitive planning, thereby significantly reducing computing and bandwidth costs while achieving high-precision, long-term reasoning for complex events. Summary of the Invention

[0008] This invention addresses the challenge posed by the explosive growth of unstructured, massive video data and the severe limitation of edge computing resources. The existing paradigm of "full-scale perception followed by passive reasoning" faces computational bottlenecks, making it difficult to achieve accurate retrieval and logical inference of sparse and complex semantic events in long-term video streams while ensuring user privacy and low latency. This invention proposes an adaptive video semantic analysis method and system based on a cognitive feedback loop. By constructing a large language model-driven iterative "reasoning-perception-verification" loop, and guided by proactive cognitive natural language queries, this invention achieves non-traversal, accurate retrieval of long-term video streams.

[0009] In a first aspect, the present invention provides an adaptive video semantic analysis method based on cognitive feedback closed loop, comprising the following steps:

[0010] Multimodal intent deconstruction and initialization: Receive query commands and long temporal video streams to be analyzed, and initialize the inference state memory containing the temporal topology;

[0011] Generative temporal planning: Based on the current reasoning state memory, perform thought chain reasoning on the query instruction to generate a set of candidate time windows with the maximum expected information gain;

[0012] Entropy-reducing sparse perception: Based on the candidate time window set, a lightweight differential algorithm is used to pre-filter invalid and redundant scenes, and visual sampling and feature encoding are completed for the target time period to obtain a high-dimensional feature vector.

[0013] Fine-grained semantic verification and causal consistency evaluation: The high-dimensional feature vector is mapped to a semantic space aligned with the query command, and the causal consistency of the event is evaluated in combination with temporal logic constraints to generate a verification feedback signal;

[0014] Dynamic graph memory evolution and strategy correction: Based on the verification feedback signal, the event graph in the reasoning state memory is updated topologically. If the termination condition is not met, the updated reasoning state memory is used to drive the large language model to correct the search strategy in the next round, forming a cognitive feedback loop.

[0015] Structured spatiotemporal reasoning response: After completing the iterative optimization of the generative temporal planning, the entropy reduction sparse perception, and the dynamic graph memory evolution and policy correction, if the global confidence exceeds the threshold, a structured analysis report containing precise timestamp positioning and event logic description is generated based on the evidence accumulated from multiple rounds of perception.

[0016] Secondly, the present invention also provides an adaptive video semantic analysis system based on a cognitive feedback closed loop, characterized in that it includes:

[0017] The cloud-based cognitive hub, deployed on a cloud server, is used to perform high-order semantic planning and global reasoning.

[0018] Edge sensing nodes, deployed locally on home smart gateways or cameras, are used for sparse sampling and feature extraction.

[0019] A secure, encrypted channel is used to connect the cloud and the edge, transmitting only de-identified visual feature vectors and text commands, while the original video stream data remains local.

[0020] The beneficial effects of this invention are:

[0021] This invention achieves Pareto optimality in computational resource consumption by employing a coarse-to-fine hierarchical planning strategy and an entropy-reduction sparse sensing mechanism. It successfully reduces the time complexity of long video understanding from a linear level. Reduced to logarithmic level This on-demand computing paradigm can accurately skip more than 90% of low-information redundant frames, thereby greatly reducing computing power and energy consumption while ensuring retrieval accuracy.

[0022] This invention establishes a high-order logical reasoning capability with "causal fidelity." Unlike traditional methods that rely solely on visual similarity matching, this invention introduces counterfactual reasoning and temporal logical constraints. This enables the system to understand the causal laws and temporal dependencies behind events beyond visual appearances (such as distinguishing between "opening" and "closing," and "cause and effect"), significantly improving robustness and accuracy for complex, time-sensitive queries.

[0023] This invention constructs a privacy protection and bandwidth optimization mechanism guided by "data sovereignty," employing an advanced feature-level federated interaction architecture. Raw pixel data is strictly confined to the local edge, with only de-identified, abstracted high-dimensional feature vectors uploaded to the cloud. This design not only physically eliminates the risk of privacy breaches but also reduces uplink network bandwidth consumption by two orders of magnitude, perfectly adapting to home broadband environments.

[0024] The system is endowed with long-term sequential memory capabilities, including "lifelong cognition." Through a dynamic event evolution graph and graph pruning mechanism, the system possesses the ability to structurally memorize ultra-long video streams. It can not only associate discrete cues with extremely long spans but also automatically clean up invalid information through a forgetting mechanism, ensuring the coherence of the reasoning context and the stability of system memory usage during long-term operation. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0026] Figure 1 A schematic diagram of the overall system architecture and collaborative reasoning process of an adaptive video semantic analysis method based on cognitive feedback closed loop provided for embodiments of the present invention;

[0027] Figure 2 This is a visual diagram illustrating the specific reasoning process for a long video question-and-answer task in an embodiment of the present invention. Detailed Implementation

[0028] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. Without conflict, embodiments and features within embodiments of the invention can be combined with each other.

[0029] Example 1: Method Flow Based on Cognitive Feedback Closed Loop

[0030] like Figure 1 As shown in the figure, this application provides an adaptive video semantic analysis method based on cognitive feedback closed loop. Figure 1 This paper demonstrates the logical connections and data interaction loops between four core functional modules: Generative Temporal Planning (RTP), Entropy-reducing Sparse Perception, Fine-grained Semantic Validation (TCR), and Dynamic Graph Memory Evolution (TM). This method reconstructs the video understanding task into a multi-round probabilistic inference process. For ease of explanation, a typical long-term complex monitoring scenario is presented: a user initiates a query command to a home security system. "Please help me find out if any stranger lingered at the door for more than 5 minutes yesterday afternoon."

[0031] Furthermore, Figure 1This demonstrates the logical connections and data interaction loops between four core functional modules: Generative Temporal Planning (RTP), Entropy-reducing Sparse Sensing, Fine-grained Semantic Validation (TCR), and Dynamic Graph Memory Evolution (TM). Specifically, the methodological steps are as follows:

[0032] S1: Multimodal Intent Deconstruction and Inference Topology Initialization

[0033] Furthermore, S1 receives natural language query instructions. and the long-time video stream to be retrieved Semantic deconstruction and state initialization are performed on the query, mapping the unstructured natural language query into a logical topology containing temporal dependencies, and initializing the inference state memory, specifically as follows:

[0034] Semantic deconstruction: The cloud-based cognitive center first processes the query command. Dependency parsing was performed to extract the core semantic graph, including the subject entity "stranger"; the action predicate "lingering"; the spatial constraint "doorway"; and the temporal constraint "duration". "min". Furthermore, the elements in the semantic graph are transformed into a directed logical topology graph containing temporally dependent edges.

[0035] State initialization: Based on the logical topology, initialize the node attributes of the temporal graph. At this point, the graph nodes only contain basic metadata of the video (such as timestamp index and low-frequency keyframe summary), and do not yet contain specific semantic events.

[0036] S2: Maximum Information Gain Guided Generative Time Series Programming (RTP)

[0037] like Figure 1 As shown, the Generative Temporal Proxy (RTP) module acts as the system's "prefrontal cortex," aiming to solve the "Where to attend" problem. The S2 module invokes a pre-trained large language model as the cognitive center, remembering the query command based on the current reasoning state. Perform chain-of-thought decomposition to predict and generate a set of candidate time windows with the maximum expected information gain. Specifically, this is achieved through the following methods:

[0038] Macro-level attention locking: Leveraging the common-sense reasoning (thought chain) capabilities of the large language model, combined with the video's global metadata curves (such as audio energy spectrum or optical flow intensity curves), a low-resolution global overview map is generated. Based on this global overview map, the large language model quickly eliminates low-probability intervals (such as periods of complete blackout or silence) and locks in the macro-level attention domain with significant audio energy fluctuations. .

[0039] Micro-window generation: Within the macro-level domain of interest, the RTP module generates a specific set of candidate time windows based on the principle of maximum information gain. Its objective function, based on the principle of maximum information gain, is constructed to find the time period that minimizes epistemic uncertainty.

[0040]

[0041] in, Indicates mutual information, For time windows Observation of expected features after performing perception operations. For current memory, For the cost of sampling calculation, This is the balance coefficient.

[0042] Furthermore, the large language model predicts that the timeframe around 15:45 is most likely to contain valid evidence, generating a set of candidate time windows. .

[0043] Furthermore, S2 reduces the temporal search space of long videos from linear complexity to logarithmic complexity through macroscopic attention locking and microscopic window generation. .

[0044] S3: Entropy-reducing adaptive sparse sensing

[0045] like Figure 1 As shown, the entropy-reducing sparse perception module is located at the edge and aims to solve the "how to perceive" problem by implementing a non-uniform sampling strategy. S3 is based on a set of candidate time windows. Dynamically allocate visual perception computing power, perform discontinuous sparse sampling and feature encoding on the target time period, and obtain local visual feature vectors. Specifically, this is achieved through the following:

[0046] Visual Proxy Pre-filtering: Before initiating high-computational feature extraction, a lightweight differential algorithm is used to calculate the bottom-up saliency of the scene. If the pixel change rate is lower than the threshold between 15:40 and 15:42, the system determines it as invalid redundancy, directly fills in the "background" placeholder, and skips depth processing.

[0047] Cross-modal attention triggering: At 15:43, before performing visual sampling, the audio stream features (Audio Fingerprint) of the corresponding time period are quickly scanned; if an acoustic event related to the semantics of the query command is detected (such as "glass breaking sound" or "dog barking"), the visual sampling rate at that time point is increased. Perform instantaneous enhancement:

[0048]

[0049] in Score the audio-text alignment. This is used to assign gain weights to audio modalities, thereby enabling precise capture of audiovisual synergy.

[0050] Feature streaming: Edge nodes encode only the selected high-value frames, reducing temporal redundancy in the video stream and generating de-identified high-dimensional feature vectors. It is then transmitted to the cloud via an encrypted channel to achieve entropy reduction in data perception.

[0051] S4: Fine-grained semantic verification and causal consistency assessment

[0052] like Figure 1 As shown, the Temporal Consistency Refiner (TCR) acts as a "discriminator," addressing the "Is it grounded?" problem. Step S4 acquires the local visual feature vector... Mapping to query command In an aligned high-dimensional semantic space, the semantic matching degree is calculated, and the causal consistency of events is evaluated in conjunction with temporal logic constraints to generate a verification feedback signal. This is achieved through the following steps:

[0053] Semantic space alignment: aligning the received feature vectors Mapping to and query The cosine similarity is calculated using the isomorphic latent space.

[0054] Counterfactual Probe: For the complex action of "wandering", TCR introduces a causal intervention mechanism to construct a counterfactual hypothesis. "The target only passes through in one direction." The system calculates the impact of the observed data on the original hypothesis. Counterfactual assumptions Likelihood Ratio:

[0055]

[0056] If the feature trajectory exhibits reciprocating characteristics on the time axis, then Logically, this confirms the "wandering" behavior and eliminates false positives.

[0057] S5: Dynamic Graph Memory Evolution and Strategy Iteration

[0058] like Figure 1 As shown, the Temporal Memory module maintains the global inference state, resolving the "contextual association" problem. Step S5, based on the verification feedback signal, performs a topological update on the event graph in the inference state memory. If the termination condition is not met, the updated memory drives the large language model to correct the search strategy for the next round. Specifically, this is implemented as follows:

[0059] Graph Topology Update: Assume the initial sensing confirms someone appeared at 15:45, but the trajectory is interrupted. The system inserts "Target detected at 15:45, Status: Interrupted" as a new node into the evolution graph. .

[0060] Closed-loop strategy correction: The RTP module reads the updated map and triggers CoT (CoT) reasoning: "Since the target appeared at 15:45 and did not leave the field of view, there must be a subsequent trajectory in the nearby time period." Based on this, the RTP module corrects the search strategy for the next round and generates a new detection window. .

[0061] Adaptive graph pruning: where the evolution graph structure Nodes in It includes three state attributes: "confirmed facts," "hypotheses to be verified," and "falsified regions"; with the number of iterations... The addition of a forget gate automatically removes isolated nodes with activation values ​​consistently below a threshold or redundant branches irrelevant to the final inference path, ensuring that the memory module's memory usage remains constant when processing extremely long video streams (such as 24-hour surveillance). Rather than growing linearly over time.

[0062] S6: Structured Spatiotemporal Reasoning Response

[0063] Furthermore, after multiple rounds of "planning-perception-correction" iterations, when the global confidence level exceeds the threshold... Upon receiving the message, the system generates a final response: "Event detected. Time: Yesterday 15:52-15:58; Object: Unknown male; Behavior: Walking back and forth in the doorway area; Confidence level: 98%." Simultaneously, the system can trigger a self-calibration interaction: if the global confidence level is still below the safety threshold at the end of the maximum iteration round, the system does not force a positive response but instead generates an "interactive clarification request." This request includes currently discovered ambiguous clues and asks the user for more constraints (e.g., "I saw an object moving at 14:00, but I'm unsure if it's your dog. Do you need further confirmation of the object's color?"). If the image is blurry, the system proactively pushes a screenshot to the user and asks: "The lighting is dim; please confirm the object's characteristics."

[0064] Specifically, this application's embodiments reconstruct the computational graph of video understanding, introducing generative artificial intelligence (GenAI) as the system's "cognitive center," and constructing an active inference architecture of "generative planning - on-demand perception - causal verification." This architecture overturns the traditional linear data flow processing mode, utilizing the high-order reasoning capabilities of large language models to establish a bidirectional adaptive feedback control loop between the semantic space and pixel space. The system no longer passively processes each frame of image, but actively predicts the spatiotemporal region with the greatest information gain based on Bayesian inference principles, drives edge perception nodes to perform non-uniform sampling, and continuously corrects prior assumptions through counterfactual reasoning mechanisms, thereby approximating the distribution of real events.

[0065] Example 2: End-to-Cloud Collaborative System Architecture

[0066] This application further illustrates the physical system architecture (logical correspondence) supporting the above method in its embodiments. Figure 1 The distribution of modules in the system (specifically including:)

[0067] A. Cloud Cognitive Orchestrator: Serving as the "brain" of the system, it is deployed on a cloud server and includes:

[0068] Core component: Equipped with a large number of parameters, the Foundation LLM is responsible for performing the logical reasoning of S1, S2, S4 and S5.

[0069] Prompt Optimizer: Contains a built-in reinforcement learning agent that monitors the planning quality of the RTP in real time. If a candidate time window is detected... If the planning is too divergent or ineffective, dynamically adjust the temperature parameters of the prompt words input to the LLM to prevent the planning from diverging.

[0070] B. Edge Agile Perceiver: Serving as the "eyes" of the system, it is deployed in smart cameras or home gateways, including:

[0071] Hardware-level wake-up unit (HW Wake-up Unit): Employs an ultra-low-power MCU to monitor audio and pixel change rates around the clock, waking up the main NPU only when a significant event is detected.

[0072] Lightweight feature extractor: Runs a small edge model (such as MobileNet) and is responsible for performing sparse sampling and feature extraction in S3.

[0073] C. Privacy & Security Layer: Employs a feature-level federated interaction protocol. The raw video stream (Pixel Data) is strictly locked to local storage at the edge, with only the abstract feature vectors, which have undergone irreversible transformation, being transmitted between the cloud and the edge, thus fundamentally protecting the data sovereignty of family privacy data.

[0074] Furthermore, the aforementioned end-to-cloud collaborative system architecture reduces uplink bandwidth usage by more than 90% through feature-level transmission rather than pixel-level transmission, while also meeting privacy compliance requirements.

[0075] Example 3: How to process complex long videos in real-world scenarios

[0076] Furthermore, to illustrate how the embodiments of this application handle complex long videos in real-world scenarios, such as... Figure 2 As shown in the figure, this illustrates how the "Temporal Memory Trace" mechanism drives the model to perform jump-like perception, summary generation, and final logical inference (Final Response) across different time segments (Selected Time). This application's embodiment demonstrates how the system answers a typical causal reasoning question: "Why is there no record of the heroine's daily life from 5:00 to 7:20 in the video?"

[0077] Iterative Perception and Temporal Memory Trace: Initial Perception (Selected time: 30-180s): The system first focuses on the introductory section at the beginning of the video. Through sparse perception, the system generates a summary: "The video opens in Miami, and the blogger introduces his sister-in-law Valeria..." This step establishes the context of the character's identity.

[0078] Intermediate Reasoning and Leaps (Selected time: 190-360s): Based on the common-sense logic that "time is continuous," the cognitive center drives the perception module to continue exploring backwards. The system perceives Valeria jogging and talking in the park, but has not yet found any key clues about the afternoon time period.

[0079] Precise positioning and evidence capture (Selected time: 960-1080s): After multiple rounds of "planning-perception" iterations, the system finally locked onto the latter part of the video. Within this time window, the system captured the key semantics: "Working hours are from 5 pm to 9 pm, and the blogger was invited to attend her last class at 7:30 pm."

[0080] Logical loop and conclusion generation: Based on the above multiple pieces of evidence, the Temporal Consistency Refiner (TCR) connects the discrete clues to construct a complete causal chain: the missing 5:00-7:20 in the video is exactly within the "teaching work time (5 PM - 9 PM)" she mentioned.

[0081] Final Response Generation: (e.g.) Figure 2 As shown, the system generated a final answer with a high confidence score (ConfidenceScore: 90): "Due to the female protagonist's promise to have two teaching classes from 5:00 to 7:30...".

[0082] Example 4

[0083] This embodiment provides an adaptive video semantic analysis system based on cognitive feedback closed loop, including:

[0084] The cloud-based cognitive hub, deployed on a cloud server, is used to perform high-order semantic planning and global reasoning.

[0085] Edge sensing nodes, deployed locally on home smart gateways or cameras, are used for sparse sampling and feature extraction.

[0086] A secure, encrypted channel is used to connect the cloud and the edge, transmitting only de-identified visual feature vectors and text commands, while the original video stream data remains local.

[0087] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0088] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An adaptive video semantic analysis method based on cognitive feedback closed loop, characterized in that, Includes the following steps: Multimodal intent deconstruction and initialization: Receive query commands and long temporal video streams to be analyzed, and initialize the inference state memory containing the temporal topology; Generative temporal planning: Based on the current reasoning state memory, perform thought chain reasoning on the query instruction to generate a set of candidate time windows with the maximum expected information gain; Entropy-reducing sparse perception: Based on the candidate time window set, a lightweight differential algorithm is used to pre-filter invalid and redundant scenes, and visual sampling and feature encoding are completed for the target time period to obtain a high-dimensional feature vector. Fine-grained semantic verification and causal consistency evaluation: The high-dimensional feature vector is mapped to a semantic space aligned with the query command, and the causal consistency of the event is evaluated in combination with temporal logic constraints to generate a verification feedback signal; Dynamic graph memory evolution and strategy correction: Based on the verification feedback signal, the event graph in the reasoning state memory is updated topologically. If the termination condition is not met, the updated reasoning state memory is used to drive the large language model to correct the search strategy in the next round, forming a cognitive feedback loop. Structured spatiotemporal reasoning response: After completing the iterative optimization of the generative temporal planning, the entropy reduction sparse perception, and the dynamic graph memory evolution and policy correction, if the global confidence exceeds the threshold, a structured analysis report containing precise timestamp positioning and event logic description is generated based on the evidence accumulated from multiple rounds of perception.

2. The adaptive video semantic analysis method based on cognitive feedback closed loop according to claim 1, characterized in that, Dependency parsing is performed on the query command to extract a semantic graph consisting of subject entities, behavioral predicates, spatial constraints, and temporal constraints.

3. The adaptive video semantic analysis method based on cognitive feedback closed loop according to claim 1 or 2, characterized in that, When generating the candidate time window set with the maximum expected information gain using a hierarchical planning strategy, the following steps are performed: Global exploration: A low-resolution global overview map is generated based on video metadata. The large language model uses this overview map to quickly eliminate low-probability time periods and lock in the macro-level domain of interest. Local fine search: Within the macro-level area of ​​interest, candidate time windows are generated by applying the principle of maximum information gain.

4. The adaptive video semantic analysis method based on cognitive feedback closed loop according to claim 3, characterized in that, When evaluating the causal consistency of an event in conjunction with the aforementioned temporal logic constraints, a counterfactual hypothesis is constructed through a causal intervention mechanism, and the target behavioral logic is verified based on the likelihood ratio calculation of the observed data.

5. The adaptive video semantic analysis method based on cognitive feedback closed loop according to claim 1, characterized in that, The trigger condition for the topology update of the event graph in the reasoning state memory is: when the first round of detection finds a target at a certain moment but the trajectory is interrupted, the empty node is inserted as a new event into the dynamic evolution graph.

6. The adaptive video semantic analysis method based on cognitive feedback closed loop according to claim 5, characterized in that, When the generative time-series planning module reads the updated topology map, it triggers a thought chain reasoning: if the target appears at a certain moment and has not left the field of vision, the generative time-series planning module corrects the next round of search strategy and generates a new set of candidate time windows.

7. The adaptive video semantic analysis method based on cognitive feedback closed loop according to claim 6, characterized in that, When updating the topology of the event graph in the reasoning state memory, an adaptive graph pruning mechanism is adopted. Low-threshold isolated nodes or historical nodes that are not related to the current reasoning path are removed through forgetting gating to ensure that the memory usage remains constant.

8. The adaptive video semantic analysis method based on cognitive feedback closed loop according to claim 1, characterized in that, If the global confidence level does not reach the threshold after the iteration is completed, a self-calibration interaction is triggered.

9. An adaptive video semantic analysis system based on cognitive feedback closed loop, characterized in that, include: The cloud-based cognitive hub, deployed on a cloud server, is used to perform high-order semantic planning and global reasoning. Edge sensing nodes, deployed locally on home smart gateways or cameras, are used for sparse sampling and feature extraction; A secure, encrypted channel is used to connect the cloud and the edge, transmitting only de-identified visual feature vectors and text commands, while the original video stream data remains local.

10. The adaptive video semantic analysis system based on cognitive feedback closed loop according to claim 9, characterized in that, The cloud-based cognitive hub also includes a prompt word optimizer for real-time monitoring of the output quality of the large language model; The edge sensing node has a hardware-level event wake-up unit embedded in it, which is used to perform audio monitoring and pixel change rate calculation.

Images