Method for dynamically recognizing actions of police dogs through multi-task learning network

Abstract

The invention provides a method for dynamically identifying actions of a police dog through a multi-task learning network, which is characterized by comprising the following steps: step 1, a data acquisition module is used for acquiring video data or sensor data of the police dog in a task execution process; step 2, a preprocessing module which preprocesses the collected video data or sensor data, including but not limited to denoising, frame extraction, image enhancement and standardization processing; 3, a feature extraction module performs feature extraction on the preprocessed image data or sensor data by using a convolutional neural network to obtain spatial-temporal feature representation of police dog actions; the method has the advantages that real-time classification, posture estimation and behavior prediction of police dog actions are achieved through the multi-task learning network, the efficiency and accuracy of a recognition system are remarkably improved, and the method can be widely applied to the fields of police dog training, law enforcement sites and security and protection and has important application value.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and biological behavior recognition, and in particular to a method for dynamically recognizing police dog actions using a multi-task learning network. Background Art

[0002] With the development of artificial intelligence technology, behavior recognition technology has been widely used in many fields. In the process of police dog training and law enforcement, accurately identifying the actions and behaviors of police dogs is of great significance for evaluating training effects, monitoring law enforcement scenes, and ensuring the safety of police dogs and personnel. Traditional action recognition methods mainly rely on single-task learning, which is difficult to simultaneously handle multiple requirements such as action classification, posture estimation, and behavior prediction. Multi-task learning, as an emerging machine learning method, can handle multiple related tasks simultaneously in the same model, improving recognition efficiency and accuracy. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for dynamically identifying police dog actions using a multi-task learning network, which solves the problem that traditional action recognition methods in the prior art mainly rely on single-task learning and are difficult to simultaneously handle multiple requirements such as action classification, posture estimation, and behavior prediction.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0005] A method for dynamically identifying police dog actions using a multi-task learning network is characterized by comprising the following steps: Step 1, a data acquisition module for collecting video data or sensor data of a police dog during a task execution;

[0006] Step 2: Preprocessing module, which preprocesses the collected video data or sensor data, including but not limited to denoising, frame extraction, image enhancement and normalization;

[0007] Step 3: Feature extraction module uses convolutional neural network to extract features from pre-processed image data or sensor data to obtain spatiotemporal feature representation of the police dog's actions;

[0008] Step 4: The multi-task learning module inputs the extracted features into a multi-task learning network, which includes the following subtasks: the action classification subtask, which classifies the police dog's actions and identifies different action types, such as running, jumping, sniffing, and attacking;

[0009] Behavior prediction task: predicting the next possible action or behavior of a police dog based on the current action sequence;

[0010] Step 5: The fusion and decision module fuses the output results of each subtask in the multi-task learning module and generates the final recognition result through weighted averaging, voting mechanism or deep fusion network;

[0011] Step 6: Output module outputs the recognition results, including action type, posture information, and behavior prediction, and can be displayed in real time or stored for subsequent analysis.

[0012] As an improvement, the acquisition module includes the following equipment: a high-resolution camera for capturing dynamic video data of the police dog; an inertial measurement unit sensor for collecting acceleration and angular velocity motion data of the police dog; and a wireless transmission module for transmitting the collected data to the preprocessing module.

[0013] As an improvement, the preprocessing module adopts the following technologies: image denoising technology, including Gaussian filtering and median filtering; frame extraction technology, extracting key frames from video data; image enhancement technology, including brightness adjustment and contrast enhancement; data normalization technology, unifying different sensor data to the same scale.

[0014] As an improvement, the feature extraction module adopts the following technologies: 3D convolutional neural network, used to extract the spatiotemporal features of video data; residual network, used to deepen the network depth and alleviate the gradient vanishing problem; attention mechanism, used to highlight important features.

[0015] As an improvement, the multi-task learning module adopts the following technologies: hard parameter sharing, which is used to share the first few layers of network parameters of each subtask; soft parameter sharing, which is used to exchange and fuse parameters between subtasks; and a multi-task loss function, which is used to balance the loss contribution of each subtask.

[0016] As an improvement, the fusion and decision module adopts the following technologies: weighted averaging method, which performs weighted averaging on the output results of each subtask; voting mechanism, which votes based on the recognition results of each subtask and selects the result with the most occurrences; deep fusion network, which inputs the output results of each subtask into a deep neural network for further fusion and decision-making.

[0017] As an improvement, the output module includes the following functions: real-time display of the police dog's action type, posture information and behavior prediction; storage of recognition results in a database for subsequent analysis and training; and provision of an alarm function, which automatically triggers an alarm when a specific action or behavior is recognized.

[0018] As an improvement, the system also includes a training module for training the multi-task learning network. The training data includes labeled police dog action video data and sensor data. The training process uses the following technologies: backpropagation algorithm for updating network parameters; stochastic gradient descent method for optimizing loss function; learning rate scheduling for dynamically adjusting the learning rate.

[0019] As an improvement, the system can be applied to the following scenarios: police dog training bases, for evaluating the training effect of police dogs; law enforcement scenes, for real-time monitoring of police dog behavior; security fields, for identifying abnormal behavior and providing early warnings.

[0020] The beneficial effects of the present invention are: real-time classification, posture estimation and behavior prediction of police dog actions are achieved through a multi-task learning network, which significantly improves the efficiency and accuracy of the recognition system. This method can be widely used in police dog training, law enforcement scenes and security fields, and has important application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a method for dynamically identifying police dog actions using a multi-task learning network. Figure 1 .

[0022] Figure 2 This is a flow chart of a method for dynamically identifying police dog actions using a multi-task learning network. Figure 2 . DETAILED DESCRIPTION

[0023] To make the contents of the present invention more clearly understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Identical parts are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, and the terms "inward" and "outward" refer to directions toward or away from the geometric center of a particular component, respectively.

[0024] like Figure 1 As shown, a method for dynamically identifying police dog actions using a multi-task learning network is characterized by comprising the following steps: Step 1, a data acquisition module for collecting video data or sensor data of the police dog during the execution of the task;

[0025] Step 2: Preprocessing module, which preprocesses the collected video data or sensor data, including but not limited to denoising, frame extraction, image enhancement and normalization;

[0026] Step 3: Feature extraction module uses convolutional neural network to extract features from pre-processed image data or sensor data to obtain spatiotemporal feature representation of the police dog's actions;

[0027] Step 4: The multi-task learning module inputs the extracted features into a multi-task learning network, which includes the following subtasks: the action classification subtask, which classifies the police dog's actions and identifies different action types, such as running, jumping, sniffing, and attacking;

[0028] Behavior prediction task: predicting the next possible action or behavior of a police dog based on the current action sequence;

[0029] Step 5: The fusion and decision module fuses the output results of each subtask in the multi-task learning module and generates the final recognition result through weighted averaging, voting mechanism or deep fusion network;

[0030] Step 6: Output module outputs the recognition results, including action type, posture information, and behavior prediction, and can be displayed in real time or stored for subsequent analysis.

[0031] The acquisition module includes the following equipment: a high-resolution camera for capturing dynamic video data of the police dog; an inertial measurement unit sensor for collecting the acceleration and angular velocity motion data of the police dog; and a wireless transmission module for transmitting the collected data to the preprocessing module.

[0032] The preprocessing module uses the following technologies: image denoising technology, including Gaussian filtering and median filtering; frame extraction technology, extracting key frames from video data; image enhancement technology, including brightness adjustment and contrast enhancement; data normalization technology, unifying data from different sensors to the same scale.

[0033] The feature extraction module uses the following technologies: 3D convolutional neural network to extract spatiotemporal features of video data; residual network to increase network depth and alleviate the gradient vanishing problem; attention mechanism to highlight important features.

[0034] like Figure 2 As shown in the figure, the multi-task learning module adopts the following technologies: hard parameter sharing, which is used to share the first few layers of network parameters of each subtask; soft parameter sharing, which is used to exchange and fuse parameters between subtasks; and multi-task loss function, which is used to balance the loss contribution of each subtask.

[0035] The fusion and decision-making module uses the following technologies: weighted averaging method, which takes a weighted average of the output results of each subtask; voting mechanism, which votes based on the recognition results of each subtask and selects the result with the most occurrences; deep fusion network, which inputs the output results of each subtask into a deep neural network for further fusion and decision-making.

[0036] The output module includes the following functions: real-time display of the police dog's action type, posture information and behavior prediction; storage of recognition results in a database for subsequent analysis and training; and provision of an alarm function, which automatically triggers an alarm when a specific action or behavior is recognized.

[0037] The system also includes a training module for training the multi-task learning network. The training data includes labeled police dog action video data and sensor data. The training process uses the following techniques: backpropagation algorithm for updating network parameters; stochastic gradient descent for optimizing the loss function; learning rate scheduling for dynamically adjusting the learning rate.

[0038] The system can be applied in the following scenarios: police dog training bases, used to evaluate the training effect of police dogs; law enforcement scenes, used to monitor the behavior of police dogs in real time; security fields, used to identify abnormal behavior and provide early warnings.

[0039] During use, high-resolution cameras and inertial measurement unit sensors are used to collect video and motion data of police dogs performing tasks in real time, and the data is sent to the preprocessing module through a wireless transmission module. The collected video and sensor data are denoised, frame extracted, image enhanced and normalized to improve data quality and unify data scale. 3D convolutional neural networks, residual networks and attention mechanisms are used to extract the spatiotemporal characteristics of police dog movements from the preprocessed data, capture their dynamic changes and key details, and input the extracted features into a multi-task learning network to try to identify the police dog's action type and predict the next possible action or behavior of the police dog based on the current action sequence. The output results of each subtask are fused, and the final recognition result is generated through weighted averaging, voting mechanism or deep fusion network to ensure the accuracy and reliability of recognition. The police dog's action type, posture information and behavior prediction results are displayed in real time to facilitate user monitoring and analysis. When specific actions or abnormal behaviors are identified, an alarm is automatically triggered to prompt the user to take corresponding measures.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for dynamic recognition of police dog actions using a multi-task learning network, characterized in that: The following steps are involved: Step 1: Data acquisition module, used to collect video data or sensor data of the police dog during the execution of the task; Step 2: Preprocessing module, which preprocesses the collected video data or sensor data, including but not limited to denoising, frame extraction, image enhancement and normalization; Step 3: Feature extraction module uses convolutional neural network to extract features from pre-processed image data or sensor data to obtain spatiotemporal feature representation of the police dog's actions; Step 4: The multi-task learning module inputs the extracted features into a multi-task learning network, which includes the following subtasks: the action classification subtask, which classifies the police dog's actions and identifies different action types, such as running, jumping, sniffing, and attacking; Behavior prediction task: predicting the next possible action or behavior of a police dog based on the current action sequence; Step 5: The fusion and decision module fuses the output results of each subtask in the multi-task learning module and generates the final recognition result through weighted averaging, voting mechanism or deep fusion network; Step 6: Output module outputs the recognition results, including action type, posture information, and behavior prediction, and can be displayed in real time or stored for subsequent analysis.

2. The method for dynamic recognition of police dog actions using a multi-task learning network according to claim 1, characterized in that: The acquisition module includes the following equipment: a high-resolution camera for capturing dynamic video data of the police dog; an inertial measurement unit sensor for collecting acceleration and angular velocity motion data of the police dog; The wireless transmission module is used to transmit the collected data to the preprocessing module.

3. The method for dynamic recognition of police dog actions using a multi-task learning network according to claim 1, characterized in that: The pre-processing module uses the following technologies: image denoising technology, including Gaussian filtering and median filtering; Frame extraction technology, extracting key frames from video data; image enhancement technology, including brightness adjustment and contrast enhancement; Data normalization technology unifies data from different sensors to the same scale.

4. The method for dynamic recognition of police dog actions using a multi-task learning network according to claim 1, characterized in that: The feature extraction module uses the following technologies: 3D convolutional neural network to extract spatiotemporal features of video data; residual network to increase network depth and alleviate the gradient vanishing problem; attention mechanism to highlight important features.

5. The method for dynamic recognition of police dog actions using a multi-task learning network according to claim 1, characterized in that: The multi-task learning module uses the following techniques: hard parameter sharing, which is used to share the first few layers of network parameters for each subtask; Soft parameter sharing is used to exchange and fuse parameters between subtasks; multi-task loss function is used to balance the loss contribution of each subtask.

6. The method for dynamic recognition of police dog actions using a multi-task learning network according to claim 1, characterized in that: The fusion and decision-making module adopts the following technologies: weighted averaging method, which takes weighted average of the output results of each subtask; voting mechanism, which votes based on the recognition results of each subtask and selects the result with the most occurrences; deep fusion network, which inputs the output results of each subtask into a deep neural network for further fusion and decision-making.

7. The method for dynamic recognition of police dog actions using a multi-task learning network according to claim 1, characterized in that: The output module includes the following functions: real-time display of the action type, posture information and behavior prediction of the police dog; The recognition results are stored in the database for subsequent analysis and training; an alarm function is provided to automatically trigger an alarm when a specific action or behavior is recognized.

8. The method for dynamic recognition of police dog actions using a multi-task learning network according to claim 1, characterized in that: The system also includes a training module for training the multi-task learning network. The training data includes labeled police dog action video data and sensor data. The training process uses the following techniques: back propagation algorithm for updating network parameters; stochastic gradient descent for optimizing the loss function; Learning rate scheduling, used to dynamically adjust the learning rate.

9. The method for dynamic recognition of police dog actions using a multi-task learning network according to claim 1, characterized in that: The system can be applied in the following scenarios: police dog training bases, for evaluating the training effect of police dogs; law enforcement scenes, for real-time monitoring of the behavior of police dogs; In the security field, it is used to identify abnormal behaviors and provide early warnings.

Images