Dynamic asset identification method and system based on AI and medium

Abstract

The invention relates to the technical field of dynamic asset identification, in particular to an AI-based dynamic asset identification method and system and a medium, which realize synchronous acquisition of image, depth and speed data of road vehicles, traffic facilities and road surface articles through a camera, a laser radar and a millimeter wave radar which are arranged on a roadside communication tower. And the accuracy and the stability of dynamic asset identification are improved. And based on a graph attention mechanism, performing target detection and behavior representation on the road assets, and generating classification tags and feature vectors. Then, a dynamic asset evaluation model is constructed through reinforcement learning, real-time evaluation of asset value is realized, and identification accuracy and response speed are optimized. For unidentified assets, classification tags and feature vectors are converted into feature words through a semantic embedding model, and asset information is traversed and matched and visualized in combination with a search engine API, so that visual and accurate asset identification and display are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic asset identification, and in particular to an AI-based dynamic asset identification method, system and medium. Background Art

[0002] With the rapid development of intelligent transportation systems, dynamic asset management in road environments has gradually become a key link in improving traffic efficiency, ensuring road safety, and optimizing resource allocation. The scope of dynamic assets covers vehicles, traffic facilities, items left on the road, and personal items carried by pedestrians. The core of its management lies in the ability to identify, monitor, and value these assets in real time and accurately. However, most current road management methods rely on manual monitoring by cameras installed on communication towers on the side of the road, which has many limitations.

[0003] Manual monitoring methods are constrained by labor costs and efficiency. During peak hours or under complex road conditions, it is difficult to quickly and comprehensively cover the dynamic changes of all road assets. Secondly, traditional methods are also difficult to meet the real-time needs of intelligent transportation in terms of response speed. The dynamic asset status in the road environment is constantly changing, such as the movement of vehicles, damage to traffic facilities, and the leaving or removal of personal belongings. These changes require real-time monitoring and rapid response. However, existing management methods usually have delays, resulting in the inability to immediately assess the value of assets after they are damaged or left behind, so as to quickly take effective measures. Summary of the invention

[0004] To solve the above problems, the present invention provides an AI-based dynamic asset identification method, system and medium, which improves the asset identification accuracy and dynamic adaptability by fusing multimodal sensor data with graph attention mechanism target detection, and further strengthening learning dynamic asset evaluation and semantic search.

[0005] To achieve the above object, the technical solution adopted by the present invention is: An AI-based dynamic asset identification method includes the following steps: S1. Use cameras, lidar and millimeter-wave radar installed on roadside communication towers to synchronously collect image, depth and speed data of road pavement objects, and perform time alignment and integration to generate a spatiotemporal feature dataset; S2. Based on the spatiotemporal feature dataset, the graph attention mechanism is used to detect and characterize the behaviors of road vehicles and road objects, and generate classification labels and feature vectors of objects. S3, based on the classification labels and feature vectors of the items, a dynamic asset evaluation model is constructed and trained through reinforcement learning, and the trained asset evaluation model is used to perform real-time asset recognition of road items. If the recognition is successful, a visual output is performed; if the recognition fails, S4 is executed; S4. Perform semantic conversion on the item classification labels and feature vectors to obtain feature words, access the search engine API, traverse the asset results through the feature words, and perform visual output.

[0006] Furthermore, the time alignment and integration comprises the following steps: De-noise the collected image data, remove outliers from the lidar data using a point cloud filtering algorithm, and smooth the velocity information of the millimeter-wave radar using adaptive filtering; Temporal alignment of images, depth and speed data of road objects, and spatial calibration combined with pre-calibrated spatial parameters; The image, depth and speed data of road objects at the same time are packaged to generate a spatiotemporal feature dataset.

[0007] Further, the S2 comprises the following steps: Based on the spatiotemporal feature dataset generated by S1, a dynamic graph structure is constructed using a graph neural network, where nodes represent the initial features of road objects and edges represent the spatiotemporal associations between objects. The nodes in the dynamic graph structure are calculated through the multi-head graph attention mechanism to extract key item features and their associations, and generate updated item feature representations; The extracted feature representations of the items are classified to generate classification labels for each item, and the dynamic trajectory of the items is behaviorally characterized through time series modeling to obtain the classification labels and feature vectors of the items.

[0008] Furthermore, the construction of a dynamic asset evaluation model through reinforcement learning based on the classification labels and feature vectors of the items includes the following steps: The classification labels and feature vectors of items are used as the input states of the reinforcement learning environment, and the objective function of asset evaluation is defined as maximizing the accuracy of asset identification and the efficiency of evaluation; Set the action space to asset valuation, priority sorting, and update operations on items; A reward function is set according to the accuracy of asset identification and the duration of evaluation, wherein the accuracy of asset identification is positively correlated with the value of the reward function, and the duration of evaluation is negatively correlated with the value of the reward function; Select the evaluation action based on the input state and calculate the immediate feedback reward of the action based on the reward function; Use gradient descent to adjust network parameters based on immediate feedback rewards.

[0009] Furthermore, the formula of the dynamic asset evaluation model is as follows: ; in, For the current state The assessed value of the assets; For the current state Next action The immediate rewards received; is the discount factor; For the next state and next action The action value function of Indicates the next state Select the maximum benefit.

[0010] Furthermore, the visual output includes: Displays real-time identified road object information in a graphical interface, including object classification and asset valuation; Use different colors to identify road items and their importance.

[0011] Further, the S4 comprises the following steps: Convert classification labels and feature vectors into feature word sets through semantic embedding models; Connect the generated feature word set to the external search engine API to construct a query request; Receive query results, extract matching asset information based on the query results, and output them visually, wherein the asset information includes name, attributes, and reference value.

[0012] An AI-based dynamic asset identification system, applied to any of the above-mentioned AI-based dynamic asset identification methods, comprising a data acquisition module, a feature extraction module, an asset identification module and a traversal search module connected in sequence; The data acquisition module is used to synchronously collect image, depth and speed data of road pavement objects through cameras, laser radars and millimeter wave radars installed on roadside communication towers, and perform time alignment and integration to generate a spatiotemporal feature data set; The feature extraction module is used to perform target detection and behavior characterization of road vehicles and road objects based on the spatiotemporal feature dataset through a graph attention mechanism to generate classification labels and feature vectors of objects; The asset identification module is used to construct and train a dynamic asset evaluation model through reinforcement learning based on the classification labels and feature vectors of the objects, and to perform real-time asset identification of road objects through the trained asset evaluation model. If the identification is successful, a visual output is performed; if the identification fails, S4 is executed; The traversal search module is used to perform semantic conversion on item classification labels and feature vectors to obtain feature words, access the search engine API to traverse asset results through the feature words, and perform visual output.

[0013] A dynamic asset identification medium based on AI, wherein the medium stores computer instructions. When a computer reads the computer instructions, the computer executes an AI-based dynamic asset identification method as described in any one of the above.

[0014] The beneficial effect of the present invention is that the present invention realizes the synchronous acquisition of images, depth and speed data of road vehicles, traffic facilities and road objects through multimodal sensors (including cameras, laser radars and millimeter wave radars) installed on roadside communication towers, overcoming the problem of insufficient adaptability of single sensors in complex scenarios. The fusion of multimodal data can effectively improve the accuracy and stability of dynamic asset recognition, especially when the lighting conditions change or the target moves quickly, it can still achieve accurate target detection and behavior characterization. Further, the target detection and behavior characterization of road vehicles and road objects are performed through the graph attention mechanism, and the classification labels and feature vectors of the objects are generated. Then, a dynamic asset evaluation model is constructed through reinforcement learning to realize real-time evaluation of asset value. The model optimizes the accuracy and response speed of asset recognition during the training process, ensuring that the value calculation can be completed quickly and the visualization results can be output after the asset status change is found. In addition, a supplementary mechanism for unidentified assets is designed. The classification labels and feature vectors of unidentified objects are converted into feature word sets through the semantic embedding model, and the external search engine API is connected to traverse and match the asset information to ensure the integrity and reliability of dynamic asset management. Combined with the output function based on a graphical interface, this solution provides an intuitive and accurate display of asset information, allowing managers to quickly understand the status and value of pavement dynamic assets and optimize resource allocation and decision-making efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a step flow chart of an AI-based dynamic asset identification method in the present invention.

[0016] Figure 2 It is a flowchart of the steps of constructing a dynamic asset evaluation model through reinforcement learning based on the classification labels and feature vectors of the items in the present invention. DETAILED DESCRIPTION

[0017] See also Figure 1-Figure 2 As shown, the present invention relates to an AI-based dynamic asset identification method, comprising the following steps: S1. Use cameras, lidar and millimeter-wave radar installed on roadside communication towers to synchronously collect image, depth and speed data of road pavement objects, and perform time alignment and integration to generate a spatiotemporal feature dataset; S2. Based on the spatiotemporal feature dataset, the graph attention mechanism is used to detect and characterize the behaviors of road vehicles and road objects, and generate classification labels and feature vectors of objects. S3, based on the classification labels and feature vectors of the items, a dynamic asset evaluation model is constructed and trained through reinforcement learning, and the trained asset evaluation model is used to perform real-time asset recognition of road items. If the recognition is successful, a visual output is performed; if the recognition fails, S4 is executed; S4. Perform semantic conversion on the item classification labels and feature vectors to obtain feature words, access the search engine API, traverse the asset results through the feature words, and perform visual output.

[0018] Specifically, multimodal sensors, including high-resolution cameras, lidar, and millimeter-wave radar, are installed on communication towers on the side of urban trunk roads to monitor dynamic road assets in real time. These sensors collect image data, three-dimensional point cloud data, and speed and distance information of objects, respectively, and work together to provide comprehensive environmental perception capabilities. The camera is used to capture the appearance images of road vehicles and objects, the lidar generates its spatial geometric structure, and the millimeter-wave radar records the speed and direction of movement of the target in real time. After data collection, a time synchronization network is used to time-align the data of different sensors, and spatial calibration is performed in combination with pre-calibrated geometric parameters to eliminate the deviation between sensors. Finally, the denoised, filtered, and aligned multimodal data are integrated into a spatiotemporal feature dataset, which contains complete characterization information for each road object. The time alignment and integration includes the following steps: denoising the collected image data, removing outliers from the laser radar data through a point cloud filtering algorithm, and smoothing the speed information of the millimeter wave radar through adaptive filtering; time aligning the image, depth and speed data of road objects, and spatially calibrating them in combination with pre-calibrated spatial parameters; packaging the image, depth and speed data of road objects at the same time to generate a spatiotemporal feature data set. Based on the generated spatiotemporal feature data set, the graph attention mechanism is used to detect targets and characterize behaviors of road vehicles and road objects. By constructing a dynamic graph neural network, each object is used as a node of the graph, and the node attributes include its appearance features, spatial position and motion state, and the edges represent the spatiotemporal associations between objects. After the classification labels and feature vectors are generated, a dynamic asset evaluation model is constructed through reinforcement learning. Taking a vehicle as an example, its classification label, speed and spatial position are used as the input state of the reinforcement learning environment, and the asset value is used as the evaluation target. The model uses a deep reinforcement learning algorithm (such as PPO) to iteratively optimize the policy network and the value network to maximize the accuracy and efficiency of asset evaluation during the training process. For example, the system can assess the degree of damage to traffic facilities in real time and assign priorities to each target item so that managers can quickly develop repair or cleanup plans. If the asset is successfully identified, the system will output the evaluation results in the form of an intuitive graphical interface, including the classification, value scoring and priority ranking of the target; if the identification fails, it will enter the next stage of semantic conversion and search. For unidentified items, the system parses their classification labels and feature vectors through a semantic embedding model to generate feature words that can accurately describe the attributes of the items. For example, when an unknown item left on the road is detected, the system converts these features into a set of feature words (such as "car, triangle mark, black, polished wheel") and retrieves possible matching results through an external search engine API. After the retrieval results are screened by relevance scoring, the best matching asset information is generated, including the name, reference value and related description, and finally output to the user interface.

[0019] Further, the S2 comprises the following steps: Based on the spatiotemporal feature dataset generated by S1, a dynamic graph structure is constructed using a graph neural network, where nodes represent the initial features of road objects and edges represent the spatiotemporal associations between objects. The nodes in the dynamic graph structure are calculated through the multi-head graph attention mechanism to extract key item features and their associations, and generate updated item feature representations; The extracted feature representations of the items are classified to generate classification labels for each item, and the dynamic trajectory of the items is behaviorally characterized through time series modeling to obtain the classification labels and feature vectors of the items.

[0020] It should be noted that based on the spatiotemporal feature dataset generated by S1, it is first necessary to use the graph neural network to build a dynamic graph structure. Each road object in the spatiotemporal feature dataset is defined as a graph node, and its attributes include initial features, such as appearance features (color, texture), three-dimensional position (point cloud coordinates), speed vector, and the size of the object. The connection relationship between nodes (i.e., the edge of the graph) is established through the spatiotemporal proximity and dynamic behavior correlation of the objects. For example, in a road scene, vehicles and objects that are close and have similar movement trends will be defined as spatiotemporal node pairs, forming a connection with a higher edge weight. After the dynamic graph structure is constructed, the nodes and edges are weighted by feature calculation through the multi-head graph attention mechanism. The core of this mechanism is to assign importance weights between each node and its neighboring nodes. Specifically, for each node, the attention mechanism calculates the feature similarity between it and its neighboring nodes, and dynamically adjusts the interaction strength between nodes in combination with the edge weight. For example, in a dynamic graph, a node represents a stationary vehicle, and its neighboring nodes are vehicles moving around. The attention mechanism can highlight the impact of stationary vehicles on the local graph structure and weaken the feature interference of fast-moving vehicles, thereby extracting a more stable dynamic feature representation. The introduction of the multi-head mechanism allows multiple attention subspaces to operate in parallel, capturing the complex correlations between nodes from different angles and generating high-quality node feature update representations. Subsequently, the extracted updated node feature representations are classified. The classification process is implemented through a fully connected network (FCN), and each node is assigned to a predefined category, such as "vehicle", "obstacle" or "leftover items" in combination with the Softmax activation function. The classification labels lay the foundation for subsequent behavior characterization and asset evaluation. For example, the detected "vehicle" node will trigger a higher frequency of dynamic behavior tracking, while the "leftover items" node needs to be further evaluated for its potential impact on traffic. Finally, the dynamic trajectory of the node is characterized by time series modeling (such as LSTM or Transformer-based models). Taking the "vehicle" node as an example, its time series data includes speed changes, movement direction and location information. Through time series modeling, its future trajectory trend can be predicted, such as "constant straight driving" or "deceleration and parking". This behavior characterization not only provides rich input data for dynamic asset evaluation, but can also be used in scenarios such as traffic flow optimization and accident warning.

[0021] Furthermore, the construction of a dynamic asset evaluation model through reinforcement learning based on the classification labels and feature vectors of the items includes the following steps: The classification labels and feature vectors of items are used as the input states of the reinforcement learning environment, and the objective function of asset evaluation is defined as maximizing the accuracy of asset identification and the efficiency of evaluation; Set the action space to asset valuation, priority sorting, and update operations on items; A reward function is set according to the accuracy of asset identification and the duration of evaluation, wherein the accuracy of asset identification is positively correlated with the value of the reward function, and the duration of evaluation is negatively correlated with the value of the reward function; Select the evaluation action based on the input state and calculate the immediate feedback reward of the action based on the reward function; Use gradient descent to adjust network parameters based on immediate feedback rewards.

[0022] In some embodiments, to construct a dynamic asset valuation model by using a reinforcement learning method, it is necessary to first use the classification label and feature vector of the item as the input state of the reinforcement learning environment. The classification label provides the category information of the item, such as "vehicle", "traffic facility" or "legacy", and the feature vector contains comprehensive information such as geometric characteristics, dynamic behavior parameters and environmental relevance. These input data constitute the decision-making basis of the agent and are used to represent the current asset status. The core goal of the reinforcement learning model is to optimize the accuracy of asset identification and the efficiency of evaluation. The accuracy of asset identification is mainly reflected in the ability of the agent to correctly identify the attributes and value of the target item, while the evaluation efficiency requires that the agent can complete the value calculation and output in the shortest time. Therefore, when constructing the model, it is clear that maximizing the recognition accuracy and minimizing the evaluation time are the optimization goals. This goal is reflected in the form of a cumulative reward function to guide the agent to continuously optimize the decision-making strategy. In the definition of the action space, the operations that the agent can perform include asset valuation, priority sorting and feature updating. For example, for the identified vehicles, the agent can assign their asset value (such as the load value of the transporter), handle different vehicles according to priority (such as priority response to high-risk vehicles), and dynamically adjust the characteristics of items according to new data. The choice of action directly affects the result of asset evaluation. The agent needs to determine the best solution among multiple possible operations to maximize the cumulative reward. The design of the reward function combines the two dimensions of asset identification accuracy and evaluation efficiency. Specifically, the reward value of accuracy is proportional to the recognition accuracy of the agent. If the asset evaluation result is highly matched with the actual value, the agent will receive a positive reward; and for the evaluation time, a time penalty mechanism is introduced. The faster the response speed, the smaller the penalty for the agent, thereby motivating the model to maintain high accuracy while completing the task efficiently. Such a reward mechanism can balance the contradiction between recognition quality and evaluation timeliness, and provide a clear direction for the decision optimization of the agent. In the training process of reinforcement learning, the agent selects actions based on the current state and updates the decision strategy based on the feedback of the reward function. The training process optimizes the policy network and the value network through repeated iterations, so that the model can gradually improve its recognition and evaluation capabilities in the ever-changing road scenes. For example, when a road object is updated from a "stationary obstacle" state to a "moving vehicle" state, the agent can re-evaluate the asset based on the real-time updated feature vector and dynamically adjust the priority and processing strategy. This continuous optimization training method ensures the adaptability and stability of the model in dynamic scenarios.

[0023] Furthermore, the formula of the dynamic asset evaluation model is as follows: ; in, For the current state The assessed value of the assets; For the current state Next action The immediate rewards received; is the discount factor; For the next state and next action The action value function of Indicates the next state Select the maximum benefit.

[0024] It should be noted that the asset valuation in the current state is equal to the weighted sum of the immediate reward and the maximum possible future benefit. The immediate reward is a direct feedback for performing a specific action in the current state, reflecting the effect of the current evaluation operation, and is usually defined by measuring the accuracy of asset identification and the efficiency of evaluation. When the valuation of the asset is close to the true value, the reward value is higher; when the evaluation takes a long time or the result deviation is large, the reward value contains a certain negative penalty, thereby motivating the agent to improve accuracy while completing the task efficiently. The discount factor is an important parameter used to balance immediate benefits and future benefits, and its value range is usually between zero and one. The introduction of the discount factor reflects the weight regulation of future benefits. When the discount factor is close to one, the model pays more attention to long-term benefits, that is, the future cumulative returns have a greater impact on the current evaluation value; when the discount factor is small, the model focuses more on the immediate benefits of the current state. By setting the discount factor reasonably, the model can achieve a dynamic balance between short-term optimization and long-term benefits. Future benefits are calculated by the action value function of the next state. The action value function is a key component in reinforcement learning, which is used to evaluate the cumulative returns that may be generated by performing an action in a specific state. In the dynamic asset evaluation scenario, the action value function combines the classification label and feature vector of the asset, and calculates the potential benefit value of each action by predicting the change trend and behavior pattern of the asset in the future. By selecting the action with the largest benefit in the future state, the model can dynamically optimize the decision direction in the current state to ensure the global optimality of the overall evaluation strategy. For example, in actual applications, when the model recognizes a vehicle asset, its classification label is "vehicle" and the feature vector includes the speed, position and appearance characteristics of the vehicle. In the current state, the agent can choose to evaluate the value of the vehicle, adjust its priority or update its feature representation. The immediate reward measures whether the currently selected operation is accurate and efficient, while the discount factor and future benefit evaluation further guide the agent on how to achieve higher cumulative returns in multi-step decision-making. When the dynamic behavior of the vehicle changes in the future state (for example, from a stationary state to an accelerated movement), the model can predict the possible benefits through the action value function and adjust the strategy in real time. During the training process, the model uses a deep reinforcement learning algorithm (deep Q learning) to continuously optimize the policy network and value network. Through repeated state transitions and reward feedback, the agent gradually improves its adaptability to asset evaluation in complex environments. For example, when faced with a scenario where a large number of assets change rapidly during rush hour, the model can quickly identify higher-priority assets and allocate resources, thereby achieving efficient dynamic asset management.

[0025] Furthermore, the visual output includes: Displays real-time identified road object information in a graphical interface, including object classification and asset valuation; Use different colors to identify road items and their importance.

[0026] In some embodiments, in a graphical interface, the information of identified road objects is displayed in real time, including the classification label, location, dynamic behavior parameters and asset valuation of the objects. Classification labels are used to distinguish the types of objects, such as "vehicles", "obstacles" or "leftover objects", while asset valuations are calculated by reinforcement learning models to reflect the relative value or importance of objects. For example, a vehicle carrying dangerous goods will be given a higher valuation and displayed in the interface for managers to prioritize. In order to improve the efficiency of information transmission, the interface uses different colors to identify road objects and their importance. The importance of objects is automatically calculated by the evaluation model based on dynamic behavior parameters and valuation results. For example, ordinary vehicles may use green labels, while dangerous objects or damaged facilities that require urgent treatment may use red labels. The visual hierarchy of colors enables users to quickly identify high-priority targets and take corresponding measures. In addition, the interface also supports interactive functions, and users can view detailed information about objects in the interface by clicking or selecting them, including the classification basis of the objects, dynamic behavior trajectories and key parameters in the evaluation process. For the simultaneous presentation of multiple objects, the interface can automatically sort them according to importance or priority, placing targets that require urgent treatment in prominent positions. For example, in actual applications, when the system detects an obstacle left on a main road, its classification label is marked as "obstacle", and the valuation model assigns it a higher degree of importance for priority processing. In the interface, the item will be displayed in eye-catching red, with dynamically updated location information and possible processing suggestions. In contrast, a normally moving vehicle will be marked in green, showing only its speed and location. Through the above-mentioned visualization design, the system can effectively improve the efficiency of dynamic asset management. Managers can not only quickly obtain information on key items, but also formulate reasonable processing plans based on asset valuation and importance, and further optimize resource allocation and operational decisions. This kind of visualization output gives full play to the actual value of dynamic asset identification and evaluation results, and provides comprehensive technical and information support for intelligent transportation systems.

[0027] Further, the S4 comprises the following steps: Convert classification labels and feature vectors into feature word sets through semantic embedding models; Connect the generated feature word set to the external search engine API to construct a query request; Receive query results, extract matching asset information based on the query results, and output them visually, wherein the asset information includes name, attributes, and reference value.

[0028] Specifically, first, the classification label and feature vector of the object are processed by the semantic embedding model and converted into a set of descriptive feature words. The classification label provides the basic semantic information of the object, such as "unknown vehicle" or "leftover object", while the feature vector contains information such as the geometric shape, dynamic behavior parameters, and environmental context of the object. The semantic embedding model uses the pre-trained BERT language model to generate a set of feature words that can accurately describe the attributes of the object by jointly analyzing the classification label and the feature vector. For example, for an unidentified round metal object, its feature word set may include key descriptions such as "round", "metal", and "stationary". Subsequently, the generated feature word set is connected to the external search engine API to construct a query request. The query request consists of feature words, combined with the dynamic behavior parameters of the object, and submitted to the search engine in a structured format. The goal of the query is to retrieve asset information matching the feature words from the external database to supplement the system's cognition of the object. For example, the system may submit the feature word set "round", "metal", and "stationary" to the search engine, and the query results may include candidate assets such as "manhole cover", "wheel", or "metal barrel". After receiving the query results, the system filters and extracts the results based on the matching degree and relevance. The matching degree is calculated based on the semantic similarity between the feature word set and the asset description in the query result, which is usually measured by cosine similarity or vector distance. The filtered matching results include the name, attributes and reference value of the asset. For example, for the above query, the best matching result after filtering may be "manhole cover", whose attributes include material is metal and reference value is a specific amount. The results are organized into a standardized format for subsequent processing. Finally, the system outputs the filtered asset information through a visualization interface for user reference. The visualization content includes the name of the item, key attributes (such as material, size) and reference value, and also displays the location and dynamic behavior of the item in a graphical form. For example, when an item is identified as a "manhole cover", the interface will display the possible use, material and valuation of the manhole cover, and graphically mark the current location and status of the item.

[0029] An AI-based dynamic asset identification system, applied to any of the above-mentioned AI-based dynamic asset identification methods, comprising a data acquisition module, a feature extraction module, an asset identification module and a traversal search module connected in sequence; The data acquisition module is used to synchronously collect image, depth and speed data of road pavement objects through cameras, laser radars and millimeter wave radars installed on roadside communication towers, and perform time alignment and integration to generate a spatiotemporal feature data set; The feature extraction module is used to perform target detection and behavior characterization of road vehicles and road objects based on the spatiotemporal feature dataset through a graph attention mechanism to generate classification labels and feature vectors of objects; The asset identification module is used to construct and train a dynamic asset evaluation model through reinforcement learning based on the classification labels and feature vectors of the objects, and to perform real-time asset identification of road objects through the trained asset evaluation model. If the identification is successful, a visual output is performed; if the identification fails, S4 is executed; The traversal search module is used to perform semantic conversion on item classification labels and feature vectors to obtain feature words, access the search engine API to traverse asset results through the feature words, and perform visual output.

[0030] Specifically, the data acquisition unit consists of a high-resolution camera, a laser radar, and a millimeter-wave radar, which are responsible for collecting images, three-dimensional point clouds, and motion status information of road objects, respectively. The camera adopts a 4K resolution, has a wide dynamic range and the ability to adapt to low-light conditions, and can clearly capture the appearance details of vehicles, obstacles, and abandoned objects. The laser radar adopts a multi-beam design (such as 32 or 64 lines) to provide high-density point cloud data for accurate modeling of the spatial geometry of objects. The millimeter-wave radar operates in the 77GHz frequency band and can measure the speed and distance information of objects in real time, which is particularly suitable for the detection of fast-moving targets. The three sensors are time-aligned through a synchronizer to ensure the synchronization and consistency of the collected data. The data acquisition unit is installed as a whole on the top of the communication tower, covering multiple lanes of the road, and realizing dynamic adjustment of the viewing angle through a universal pan-tilt head. The feature extraction module is an integrated computing unit equipped with a GPU accelerator card or a dedicated deep learning inference chip for real-time processing of multimodal data transmitted from the data acquisition module. Edge computing devices such as NVIDIA Jetson platform or high-performance FPGA cards are used in hardware to support the following functions: Data preprocessing: integrated denoising filter unit to suppress noise on camera images, remove outliers on lidar point clouds, and smooth millimeter wave radar signals. Feature extraction calculation: built-in hardware acceleration unit supporting graph neural network, used to perform efficient calculation of dynamic graph construction and multi-head graph attention mechanism. This module can generate classification labels and feature vectors of items in real time, providing input for subsequent asset identification modules. An edge computing unit is deployed in the communication tower cabinet, responsible for the operation of asset identification module and traversal search module. This unit adopts server-level hardware configuration, including multi-core CPU, GPU acceleration card and large-capacity memory, supporting complex reinforcement learning algorithms and deep learning reasoning. Asset identification hardware: equipped with hardware accelerator of reinforcement learning model, realizes asset identification and value assessment through analysis and prediction of real-time data. Real-time processing capability: supports large-scale parallel computing, capable of processing multi-modal input and high-frequency data streams. Built-in cooling system ensures the stability of equipment under high load.

[0031] The present invention also includes an AI-based dynamic asset identification medium, which stores computer instructions. When a computer reads the computer instructions, the computer executes an AI-based dynamic asset identification method as described in any of the above items.

[0032] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A dynamic asset identification method based on AI, characterized in that: The following steps are involved: S1. Use cameras, lidar and millimeter-wave radar installed on roadside communication towers to synchronously collect image, depth and speed data of road pavement objects, and perform time alignment and integration to generate a spatiotemporal feature dataset; S2. Based on the spatiotemporal feature dataset, the graph attention mechanism is used to detect and characterize the behaviors of road vehicles and road objects, and generate classification labels and feature vectors of objects. S3. Based on the classification labels and feature vectors of the items, a dynamic asset evaluation model is constructed and trained through reinforcement learning. The trained asset evaluation model is used to perform real-time asset recognition of road items. If the recognition is successful, a visual output is performed. If the recognition fails, execute S4; S4. Perform semantic conversion on the item classification labels and feature vectors to obtain feature words, access the search engine API, traverse the asset results through the feature words, and perform visual output.

2. The AI-based dynamic asset identification method according to claim 1, characterized in that: The time alignment and integration comprises the following steps: De-noise the collected image data, remove outliers from the lidar data using a point cloud filtering algorithm, and smooth the velocity information of the millimeter-wave radar using adaptive filtering; Temporal alignment of images, depth and speed data of road objects, and spatial calibration combined with pre-calibrated spatial parameters; The image, depth and speed data of road objects at the same time are packaged to generate a spatiotemporal feature dataset.

3. The AI-based dynamic asset identification method according to claim 1, characterized in that: The S2 comprises the following steps: Based on the spatiotemporal feature dataset generated by S1, a dynamic graph structure is constructed using a graph neural network, where nodes represent the initial features of road objects and edges represent the spatiotemporal associations between objects. The nodes in the dynamic graph structure are calculated through the multi-head graph attention mechanism to extract key item features and their associations, and generate updated item feature representations; The extracted feature representations of the items are classified to generate classification labels for each item, and the dynamic trajectory of the items is behaviorally characterized through time series modeling to obtain the classification labels and feature vectors of the items.

4. The AI-based dynamic asset identification method according to claim 3, characterized in that: The method of constructing a dynamic asset evaluation model through reinforcement learning based on the classification labels and feature vectors of the items includes the following steps: The classification labels and feature vectors of items are used as the input states of the reinforcement learning environment, and the objective function of asset evaluation is defined as maximizing the accuracy of asset identification and the efficiency of evaluation; Set the action space to asset valuation, priority sorting, and update operations on items; A reward function is set according to the accuracy of asset identification and the duration of evaluation, wherein the accuracy of asset identification is positively correlated with the value of the reward function, and the duration of evaluation is negatively correlated with the value of the reward function; Select the evaluation action based on the input state and calculate the immediate feedback reward of the action based on the reward function; Use gradient descent to adjust network parameters based on immediate feedback rewards.

5. The AI-based dynamic asset identification method according to claim 4, characterized in that: The formula of the dynamic asset valuation model is as follows: ； in, For the current state The assessed value of the assets; For the current state Next action The immediate rewards received; is the discount factor; For the next state and next action The action value function of Indicates the next state Select the maximum benefit.

6. The AI-based dynamic asset identification method according to claim 1, characterized in that: The visual output includes: Displays real-time identified road object information in a graphical interface, including object classification and asset valuation; Use different colors to identify road items and their importance.

7. The AI-based dynamic asset identification method according to claim 1, characterized in that: The S4 comprises the following steps: Convert classification labels and feature vectors into feature word sets through semantic embedding models; Connect the generated feature word set to the external search engine API to construct a query request; Receive query results, extract matching asset information based on the query results, and output them visually, wherein the asset information includes name, attributes, and reference value.

8. An AI-based dynamic asset identification system, applied to an AI-based dynamic asset identification method according to any one of claims 1 to 7, characterized in that: It includes a data acquisition module, a feature extraction module, an asset identification module and a traversal search module which are connected in sequence; The data acquisition module is used to synchronously collect image, depth and speed data of road pavement objects through cameras, laser radars and millimeter wave radars installed on roadside communication towers, and perform time alignment and integration to generate a spatiotemporal feature data set; The feature extraction module is used to perform target detection and behavior characterization of road vehicles and road objects based on the spatiotemporal feature dataset through a graph attention mechanism to generate classification labels and feature vectors of objects; The asset identification module is used to construct and train a dynamic asset evaluation model through reinforcement learning based on the classification labels and feature vectors of the objects, and to perform real-time asset identification of road objects through the trained asset evaluation model. If the identification is successful, a visual output is performed; if the identification fails, S4 is executed; The traversal search module is used to perform semantic conversion on item classification labels and feature vectors to obtain feature words, access the search engine API to traverse asset results through the feature words, and perform visual output.

9. An AI-based dynamic asset identification medium, wherein the medium stores computer instructions. When a computer reads the computer instructions, the computer executes an AI-based dynamic asset identification method as described in any one of claims 1 to 7.

Images