Abnormal behavior identification method based on hotel monitoring and dynamic inspection

Abstract

The invention relates to the field of safety protection and emergency management, and discloses an abnormal behavior identification method based on hotel monitoring and dynamic inspection, which comprises the following steps: modeling a hotel area, dividing a monitoring range into a plurality of nodes, and calculating the risk weight of each node based on historical behavior data and real-time collected data; based on the risk weight, generating an optimal inspection path of the inspection robot through a path planning optimization algorithm; and in a multi-robot cooperation scene, the inspection path is dynamically adjusted based on the game optimization model, and inspection resource conflicts are avoided. Through cooperation of the inspection robot, the fixed camera, the path planning and the game optimization module, efficient inspection path planning and abnormal behavior recognition are realized, path selection is optimized, blind areas are avoided, the accuracy of abnormal behavior recognition is improved by adopting a spatial-temporal feature extraction and deep learning model, the game optimization model coordinates cooperation of multiple robots, and the accuracy of abnormal behavior recognition is improved. And path conflicts are reduced, and the system adaptability is enhanced through intelligent path planning and dynamic adjustment.

Description

Technical Field

[0001] The present invention relates to the technical field of safety prevention and emergency management, and in particular to an abnormal behavior recognition method based on hotel monitoring and dynamic inspection. Background Art

[0002] With the development of society, the safety of public places has received increasing attention, especially in crowded places such as hotels. Existing security measures mostly rely on manual patrols and fixed camera video playback. Although this post-processing method can provide a certain degree of security, it is inefficient due to the inability to intervene in real time, and it is easy to miss the best time for intervention. In addition, fixed cameras usually only cover specific areas, have blind spots, and are limited by viewing angles and environmental changes. They may not be able to capture all abnormal behaviors, resulting in false alarms and missed alarms.

[0003] In recent years, the progress of computer vision technology and deep learning algorithms has made it possible to automatically detect abnormal behavior in video surveillance. Compared with traditional image processing methods, this abnormal behavior recognition system based on deep learning can not only improve the timeliness of detection, but also reduce the dependence on manual monitoring and reduce labor costs. However, the existing abnormal behavior recognition system relies on static cameras to capture images and uses image processing algorithms to identify preset behavior patterns. Although this method can cover a larger monitoring area, it also has obvious limitations. First, static cameras have a fixed shooting angle, which makes it difficult to cover hidden locations or capture abnormal behaviors in specific situations, which can easily lead to omissions. Secondly, traditional image processing algorithms can usually only recognize some simple and clear actions, and have poor adaptability to complex, changing environments and dynamic behaviors.

[0004] In addition, most existing systems lack active inspection mechanisms, cannot flexibly adjust inspection routes, and fail to make full use of dynamic equipment such as inspection robots to supplement the blind spots of fixed cameras, resulting in the inability to conduct comprehensive and detailed real-time scanning of the space, thereby reducing the overall level of security. Therefore, although existing technologies can provide monitoring support in some cases, they are still insufficient in terms of efficiency, comprehensiveness, and real-time response, and more advanced technical solutions are urgently needed to improve the intelligence level and adaptability of monitoring systems. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention provides an abnormal behavior identification method based on hotel monitoring and dynamic inspection, which realizes efficient inspection path planning and abnormal behavior identification, improves the flexibility and adaptability of the system, and solves the problems of insufficient dynamic response, path conflict and recognition accuracy of traditional monitoring systems. At the same time, through deep learning models and intelligent path optimization mechanisms, abnormal behaviors in complex scenarios can be captured in real time and inspection strategies can be dynamically adjusted.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: an abnormal behavior identification method based on hotel monitoring and dynamic inspection, comprising the following steps: Model the hotel area, divide the monitoring range into multiple nodes, and calculate the risk weight of each node based on historical behavior data and real-time collected data; Based on the risk weight, the optimal inspection path of the inspection robot is generated through the path planning optimization algorithm; In multi-robot collaboration scenarios, the inspection path is dynamically adjusted based on the game optimization model to avoid inspection resource conflicts; The inspection robot collects images and environmental data in real time along the optimal path, and identifies abnormal behaviors based on spatiotemporal feature extraction and behavior prediction models; The identified abnormal behavior data is uploaded to the cloud, and the node risk weights and behavior prediction model parameters are dynamically updated.

[0007] Preferably, the calculation of the risk weight of each node comprises the following steps: Based on historical inspection data, the frequency of abnormal behavior of each node is counted to obtain the historical risk weight; Based on the real-time data collected by the inspection robot and fixed cameras, the risk value of the current node is predicted in combination with the abnormal behavior recognition model to obtain the real-time risk weight; The historical risk weights and real-time risk weights are integrated in a predetermined ratio to obtain a comprehensive risk weight.

[0008] Preferably, the calculation formula of the comprehensive risk weight is: Among them, α and β are predetermined weight coefficients, which represent the weight ratio of historical data and real-time data in the comprehensive risk weight, and satisfy α+β=1. Represents the historical risk weight of the node, which is used to reflect the frequency of abnormal behavior of the node in the historical behavior data. Indicates the real-time risk weight of the node, which is used to reflect the probability of abnormal behavior of the node at the current moment.

[0009] Preferably, the steps of the path planning optimization algorithm include: The monitoring nodes in the hotel environment are constructed into a graph structure, where the nodes represent the inspection areas and the edge weights represent the inspection time; Setting multi-objective optimization goals, including minimizing the total time of the inspection path, minimizing the risk value of uncovered high-risk nodes, and minimizing redundant coverage of the path; Use non-dominated sorting genetic algorithm to solve and generate Pareto solution set according to non-dominated sorting; The optimal solution that balances path time, risk coverage and redundancy is selected from the Pareto solution set as the inspection path.

[0010] Preferably, the solution using a non-dominated sorting genetic algorithm comprises the following steps: Randomly generate several initial inspection paths and calculate the evaluation value of each path on the multi-objective function; Performance on all targets,classifies the population into several non-dominated classes; The crowding distance is calculated for each individual to evaluate population diversity; Select individuals with higher non-dominant ranks and larger crowding distances from the population as parents; Perform a crossover operation on the selected parent generation to generate offspring, and perform a mutation operation on the offspring; Merge the parent and child generations and select a new population based on non-dominated sorting and crowding distance; Repeat the above steps until the preset number of iterations and optimization conditions are met.

[0011] Preferably, the dynamic adjustment of the inspection path based on the game optimization model specifically includes the following steps: A non-cooperative game model of patrol robot collaboration is constructed, where each robot calculates its own path benefit based on a benefit function, which includes the total risk benefit of path coverage and the penalty cost of path conflict; The optimal hybrid strategy for each robot to select a path is solved by iteratively updating the probability distribution of each robot's path selection, and the probability distribution of the path selection is dynamically optimized by the benefit gradient method; After determining the path of each robot, the optimal path combination of all inspection robots under the game model is obtained by solving the Nash equilibrium, which avoids path conflicts and ensures priority coverage of high-risk areas.

[0012] Preferably, the Nash equilibrium satisfies the following conditions: in, For robot i in strategy and other robot strategies The following income, Robot i in other strategies S i ' and other robot strategies The income under S i is the path selection set of robot i, S -i The set selected for all other robot paths.

[0013] Preferably, the abnormal behavior identification based on spatiotemporal feature extraction and behavior prediction model includes the following steps: extracting the time features of the image data collected by the inspection robot, and using the time dimension attention mechanism to capture the dynamic changes of abnormal behavior; Extract the spatial features of image data and use the spatial dimension attention mechanism to determine the area where abnormal behavior occurs; The temporal and spatial features are processed jointly to obtain the comprehensive feature weights; The comprehensive feature weights are input into the long short-term memory network model to capture the long-term dependencies of the behavior sequence; Based on the recognition results output by the model, the behavior is classified to determine whether it is abnormal behavior.

[0014] Preferably, the abnormal behavior identification result is used to dynamically update the node risk weight and behavior prediction model parameters, including the following steps: Based on the abnormal behavior identification results, the real-time risk weight of the corresponding node is adjusted; Upload the behavior data and recognition results collected during the inspection process to the cloud to generate a new behavior sample set; Retrain the behavior prediction model with a new behavior sample set and update the model parameters; The updated risk weights and model parameters are sent to the path planning module and abnormal behavior identification module for optimization of the next round of inspection tasks.

[0015] The present invention also provides an abnormal behavior identification system based on hotel monitoring and dynamic inspection, comprising: Inspection robot: used to perform inspection tasks, equipped with a variety of sensors, cameras and behavior recognition modules, to collect environmental data and video data, and analyze abnormal behavior in real time; Fixed surveillance cameras: deployed at multiple nodes within the hotel to collect static surveillance images and provide auxiliary inspection data; Path planning module: used to receive node risk weight information and generate the optimal inspection path of the inspection robot based on a multi-objective optimization algorithm; Collaboration optimization module: used to optimize the multi-robot collaboration strategy based on the game model and avoid inspection conflicts through dynamic path adjustment; Abnormal behavior identification module: used to identify abnormal behaviors of the collected inspection data based on the spatiotemporal feature extraction mechanism and the long short-term memory network model, and generate abnormal behavior classification results; Data processing and feedback module: used to receive inspection results and identification data, dynamically adjust node risk weights and train and update behavior prediction models to complete system parameter optimization; Cloud server: used to store inspection data, recognition results and historical behavior samples, and supports remote updating and deployment of path planning optimization and behavior prediction models.

[0016] The present invention provides an abnormal behavior identification method based on hotel monitoring and dynamic inspection. It has the following beneficial effects: 1. The present invention adopts a multi-module collaborative technical solution including inspection robots, fixed cameras, path planning modules, collaborative optimization modules, etc., to achieve efficient inspection path planning and abnormal behavior recognition. Compared with the single inspection or monitoring system in the prior art, the present invention optimizes path selection and avoids duplicate coverage or blind spots through real-time data sharing and feedback mechanisms.

[0017] 2. The present invention adopts the technical solution of spatiotemporal feature extraction and behavior prediction deep learning model, combined with convolutional neural network (CNN) and long short-term memory network (LSTM), to achieve higher abnormal behavior recognition accuracy and timeliness. Compared with the abnormal behavior recognition system based on simple pattern matching in the prior art, the present invention can capture abnormal behaviors in complex scenes in real time, and improves the recognition ability of various complex behaviors through deep learning models.

[0018] 3. The present invention optimizes the path selection in multi-robot collaboration through the technical solution of the game optimization model, achieving the technical effect of reducing path conflicts between robots and improving resource utilization. Compared with the lack of effective path coordination technical solutions in the prior art multi-robot system, the present invention ensures that each robot can flexibly adjust the path according to real-time data through the game optimization model, effectively avoiding the problem of resource waste and low inspection efficiency.

[0019] 4. The present invention achieves the technical effect of high flexibility and dynamic adjustment through intelligent path planning and game optimization mechanism. Compared with the existing technology that relies on fixed paths and simple inspection routes, the present invention can adjust the inspection route according to real-time data and risk assessment results, so that the system can flexibly respond to dynamic changes in the hotel environment, such as changes in passenger flow, regional safety hazards, etc., and improve the overall adaptability and response capabilities of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the present invention; Figure 2 It is a system framework diagram of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Please see attached Figure 1 The embodiment of the present invention provides an abnormal behavior identification method based on hotel monitoring and dynamic inspection, comprising the following steps: S1. Model the hotel area, divide the monitoring range into multiple nodes, and calculate the risk weight of each node based on historical behavior data and real-time collected data; S2. Based on the risk weight, the optimal inspection path of the inspection robot is generated through the path planning optimization algorithm; S3. In the scenario of multi-robot collaboration, the inspection path is dynamically adjusted based on the game optimization model to avoid inspection resource conflicts; S4, the inspection robot collects images and environmental data in real time along the optimal path, and identifies abnormal behaviors based on spatiotemporal feature extraction and behavior prediction models; S5. Upload the identified abnormal behavior data to the cloud, and dynamically update the node risk weight and behavior prediction model parameters.

[0023] Specifically, in the implementation process of the present invention, by scientifically modeling the hotel area, combining historical inspection data and real-time data collection to calculate the risk weight of the node, an important basis is provided for the planning of dynamic inspection paths and the efficient identification of abnormal behaviors. The calculation of node risk weights is both the starting step of the present invention and an important basis for path optimization and behavior identification. By constructing a node risk weight evaluation system, inspection resources can be scientifically allocated, high-risk areas can be covered preferentially, and timely response capabilities can be provided for emergencies in a dynamic environment.

[0024] In step S1, the modeling of the hotel area is based on a graph structure model, where: Nodes represent areas that need to be inspected, such as key locations such as the hotel's front desk, corridors, elevator entrances, and exits; edges represent the relationships between nodes, such as physical distance, inspection time, or communication cost of a path; it should be noted that the risk weight of each node is composed of historical risk weight and real-time risk weight, and the comprehensive risk weight of the node is calculated by weighted fusion.

[0025] In a possible implementation, the historical risk weight is calculated based on the statistical results of abnormal behaviors in historical inspection records. For example, the historical risk weight of a node can be expressed as: Among them, N abnormal,i Indicates the total number of abnormal behaviors that occurred in the node in the historical data: N total,i Indicates the total number of times the node has been inspected in historical data. is the historical risk weight of the node, which is used to reflect the frequency of abnormal behavior of the node in historical inspection records.

[0026] Specifically, this calculation method can quantify the probability of abnormal behavior of a node in long-term operation. For example, in some embodiments, the node located at the front desk of a hotel has a high number of abnormal behaviors due to dense traffic, so the historical risk weight of the node is also correspondingly high, so it is given priority in path planning. The historical risk weight can also be corrected in combination with the long-term inspection coverage rate, thereby improving the applicability of historical data.

[0027] The real-time risk weight is calculated based on the real-time data collected by the inspection robot or fixed camera through spatiotemporal feature extraction and behavior recognition model. Specifically, this part of the weight reflects the probability of abnormal behavior at the node at the current moment.

[0028] In a possible implementation, the steps of calculating the real-time risk weight include: Extract temporal and spatial features from image data collected by inspection robots or fixed cameras; Use the time dimension attention mechanism to capture the dynamic changes of abnormal behaviors, such as the duration and frequency of actions; Use the spatial dimension attention mechanism to focus on key areas in the picture, such as gathering points and locations where abnormal actions occur; input temporal and spatial features into the long short-term memory network (LSTM) model to evaluate the probability of abnormal behavior of the current node; The specific expression of real-time risk weight is: in, is the real-time risk weight of the node, which is used to reflect the probability of abnormal behavior of the node at the current moment. abnormal,i The predicted probability of abnormal behavior of the node is obtained based on the spatiotemporal feature extraction and behavior recognition model. The present invention calculates the comprehensive risk weight R of the node by integrating the historical risk weight and the real-time risk weight. i The specific formula is: Among them, α and β are predetermined weight coefficients, which represent the weight ratio of historical data and real-time data in the comprehensive risk weight, and satisfy α+β=1. Represents the historical risk weight of the node, which is used to reflect the frequency of abnormal behavior of the node in the historical behavior data. Indicates the real-time risk weight of the node, which is used to reflect the probability of abnormal behavior of the node at the current moment.

[0029] In some embodiments, the present invention can be combined with cloud computing capabilities to periodically update node risk weights, thereby improving the adaptability of the system. For example: through global behavior data analysis, the values ​​of α and β are dynamically adjusted to make the weight distribution more consistent with the current inspection environment. According to the historical trends of different nodes, the weights of long-term low-risk nodes are reduced to save inspection resources, and the edge weights are corrected so that the inspection path can cover high-risk areas more efficiently. Exemplarily, when a node has not exhibited abnormal behavior for a long time, its comprehensive risk weight can be gradually reduced by adjusting α and β, thereby optimizing the rationality of inspection resource allocation.

[0030] For step S2, step S2 is based on the node comprehensive risk weight R calculated in step S1. i , combined with the path planning optimization algorithm, the optimal inspection path is generated for the inspection robot. The path planning aims to maximize the coverage of high-risk areas, minimize the inspection time and reduce the path redundancy. The present invention adopts a multi-objective optimization strategy, comprehensively considers multiple planning objectives and constraints, and uses a non-dominated sorting genetic algorithm (NSGA-II) to generate a Pareto optimal solution set, and selects the optimal path suitable for the inspection needs.

[0031] The node comprehensive risk weight provides a reasonable risk priority for path planning. The optimization of path planning results directly affects the efficiency of inspection tasks and the coverage of high-risk areas. For example, the path planning of the present invention is suitable for single-robot inspection tasks, and has strong scalability and can support multi-robot collaborative inspections.

[0032] In this embodiment, the hotel monitoring area is modeled as a directed weighted graph G = (V, E), where: Node set V = {v 1 ,v 2 ,…,v n} represents the monitoring area that needs to be inspected. For example, the hotel front desk, elevator entrance, and corridor; edge set E = {e ij} represents the inspection path between nodes, and its weight w ij Usually from node v i To node v j inspection time or physical distance.

[0033] The goal of path planning is to calculate the comprehensive risk weight R of the graph G and the nodes. i , generating an optimal path P to meet the requirements of covering high-risk areas, optimizing inspection time and resource utilization.

[0034] The formula for minimizing the total path time is: Among them, f1 (P) is the total inspection time of path P, P is the inspection path, which consists of several nodes and edges, and (i,j) is the number of nodes from node v in the path. i To node v j The edge of w ij For slave node v i To node v j inspection time or physical distance.

[0035] Minimize the risk value calculation of uncovered high-risk nodes: Among them, f 2 (P) is the sum of the comprehensive risk values ​​of the nodes not covered by path P, v i Monitoring nodes that need to be inspected, R i v i The comprehensive risk weight of the node is calculated by the formula in step S1. Represents node v i Not included in path P.

[0036] Minimize path redundancy calculation: Among them, f 3 (P) is the redundant coverage number of path P, δ ij is the identification variable of whether the path (i, j) is visited repeatedly, δ ij =1 means that the path is visited repeatedly, δ ij =0 means it has not been visited repeatedly.

[0037] The constraints for path planning are: Among them, x ij Is the identification variable of whether the path contains edge (i, j), x ij =1 means included, x ij =0 means not included, represents the sum of all nodes j, It means that the constraint must be satisfied for all nodes i, and n represents the total number of nodes.

[0038] This constraint ensures that each node is visited once in path planning and the path is a simple loop, meeting the inspection requirements.

[0039] Limitation of high-risk node coverage: Where ρ is the coverage rate of high-risk nodes in path P, P is the inspection path, and v i Monitoring nodes that need to be inspected, Ri v i The comprehensive risk weight of the node, θ is the minimum threshold of high-risk node coverage, which is used to ensure the coverage effect of the path in the high-risk area.

[0040] The constraint condition sets a minimum coverage rate θ to ensure that high-risk areas are fully inspected.

[0041] The solution process of the non-dominated sorting genetic algorithm (NSGA-II) includes: Randomly generate N initial inspection paths to form the initial population. Each path must meet the above constraints. For each path P in the population, calculate the objective function value separately; According to the objective function value, the population is divided into multiple non-dominated levels. The better individuals are retained and optimized according to the non-dominated sorting; The crowding distance is calculated for each individual to measure the diversity of the population. The calculation formula is: Among them, d i is the crowding distance of individual i, m is the number of optimization targets, and For individual i in target f k The objective function value of the adjacent individuals on ; Perform crossover and mutation operations on the offspring paths, and generate new path individuals through genetic algorithm operations; Merge the parent and child generations and select the new population based on non-dominated sorting and crowding distance; The above process is repeated until the convergence condition is reached (for example, the number of iterations or the Pareto solution set is stable).

[0042] The present invention uses a path planning optimization algorithm and a non-dominated sorting genetic algorithm (NSGA-II) to efficiently solve multi-objective optimization problems and generates an inspection path with the shortest path time, high-risk coverage priority and lowest redundancy.

[0043] In step S3, the present invention uses a method based on spatiotemporal feature extraction and behavior prediction model to analyze the image data collected by the inspection robot to identify possible abnormal behaviors. These abnormal behaviors include but are not limited to smoking in non-smoking areas, fighting and other malicious incidents. The spatiotemporal feature extraction model combines information in time and space dimensions, so that more accurate behavioral features can be extracted in different scenarios. By further utilizing deep learning models such as long short-term memory networks (LSTMs), combined with historical behavior data and real-time data, the accuracy and timeliness of abnormal behavior identification can be effectively improved.

[0044] It should be noted that the core of step S3 is to extract the spatiotemporal information obtained from the video data during the inspection process, analyze the long-term dependencies of the behavior sequence in combination with the deep learning model, and finally achieve accurate detection of abnormal behavior. In some embodiments, the spatiotemporal feature extraction and behavior prediction model will adopt a more complex neural network architecture, such as a deep learning model combining a convolutional neural network (CNN) and a recurrent neural network (RNN), to further improve the recognition accuracy.

[0045] First, the image data collected by the inspection robot is used to extract spatial features. These image data contain information about people, objects, and the environment in the scene, so they need to be processed by computer vision technology to extract meaningful spatial features from the image. For example, a convolutional neural network (CNN) is used to process each frame of the image to extract features related to objects, people's postures, and activities.

[0046] In this embodiment, the spatiotemporal feature extraction process includes the following steps: Use a convolutional neural network (CNN) to process each frame of the image and extract spatial information about people, objects, and actions. For example, identify the people in the image and their behavior patterns.

[0047] For an image I(t), its spatial feature F s (t) can be expressed as: F s (t) = CNN(I(t)) Among them, F s (t) represents the spatial features at time t, and CNN(I(t)) represents the spatial features extracted from image I(t) through a convolutional neural network.

[0048] Since the inspection task needs to consider the continuity and long-term dependency in the time series, the time dimension attention mechanism is adopted to capture the dynamic changes of abnormal behavior by modeling the image sequence. t (t) can be modeled by a long short-term memory network (LSTM): F t (t) = LSTM(F s (1),F s (2),…,F s (t)) Among them, F t (t) represents the spatial features at time t, and the time-dependent information extracted by the LSTM network, F s (1),F s (2),…,F s (t) represents the spatial characteristics at time 1, 2, …, t in the time series.

[0049] Through the process of spatiotemporal feature extraction, the robot can obtain information related to space and time at each moment, and perform joint analysis through a deep learning model to capture the changing pattern of behavior. Based on the spatiotemporal feature extraction, a behavior prediction model is used in this embodiment to further determine whether there is abnormal behavior. The behavior prediction model usually consists of two parts: one is the extraction of spatiotemporal features, and the other is the identification of abnormal behavior based on the extracted spatiotemporal features.

[0050] The time feature F t (t) and spatial characteristics F s (t) are processed together to obtain a comprehensive feature representation F(t), which represents the overall behavior state at the current time t: F(t)=f(F s (t),F t (t)) Among them, F(t) represents the comprehensive features at time t, and f(·) represents the function that combines spatial and temporal features.

[0051] The joint feature F(t) is input into the classification model to classify abnormal behavior. Usually, this classification model can use a neural network-based classifier, such as a fully connected layer (FC) or a support vector machine (SVM). The final output classification result will indicate whether the current behavior is abnormal.

[0052] Y(t)=Predict(F(t)) Where Y(t) represents the behavior category recognized by the robot at time t. If it is an abnormal behavior, it is "1", otherwise it is "0". Predict(·) represents the operation predicted by the classifier.

[0053] Once the behavior prediction model determines that the behavior at a certain moment is abnormal (i.e., Y(t) = 1), the system will upload the abnormal behavior information to the cloud in real time. The uploaded data includes but is not limited to: the specific time t when the abnormal behavior occurred, the area or node v where the abnormal behavior occurred, and the time when the abnormal behavior occurred. i , specific types of abnormal behavior.

[0054] The game optimization model is a mathematical model commonly used for decision-making and resource allocation problems in multi-agent systems. The overall optimization of the system is achieved by simulating the interaction and competition between multiple decision makers (or "agents"). Each agent makes decisions based on its own goals (usually maximizing utility), and these decisions will affect each other in the system.

[0055] In the application scenario of the present invention, the game optimization model is introduced for dynamic collaboration and path optimization between inspection robots. Specifically, in the multi-robot collaboration scenario, the inspection robots need to adjust the path according to real-time data and avoid path conflicts while ensuring priority coverage of high-risk areas. The game optimization model simulates the path selection and resource competition between robots, so that each robot can find an optimal balance between collaboration and competition.

[0056] The game optimization model usually consists of the following parts: In the present invention, each inspection robot is an intelligent agent. Each intelligent agent needs to choose an optimal path to meet its own inspection goals while avoiding conflicts with the paths of other robots; A set of strategies that each agent can choose. In this invention, strategy refers to the path selection of the inspection robot. The robot can choose multiple different paths, and the size of the strategy space is related to the layout of the hotel and the number of robots; The utility function is a function that evaluates the "gains" or "losses" of an intelligent agent (inspection robot) after choosing a certain strategy (path). In the present invention, the utility function of the robot's path selection takes into account two main factors: the safety benefits brought by covering high-risk areas, and the penalty costs that will be incurred if two robots choose the same or conflicting paths; The goal of each agent is to maximize its utility function, that is, to choose a path that covers the most high-risk areas while minimizing the cost caused by path conflicts; The solution to the game refers to the strategy chosen by each agent, so that given the strategies of other agents, no agent can obtain a better result by unilaterally changing its own strategy. That is, each robot chooses the optimal path under the constraints of the strategies of other robots and reaches an equilibrium state.

[0057] The core goal of the game optimization model is to find the Nash equilibrium. In the Nash equilibrium, each robot will choose the optimal path, so that when the path selection of other robots remains unchanged, it cannot gain greater benefits by changing its own path selection.

[0058] A Nash equilibrium satisfies the following conditions: in, For robot i in strategy and other robot strategies The following income, Robot i in other strategies S i ' and other robot strategies The income under S iis the path selection set of robot i, S -i The set selected for all other robot paths.

[0059] The game optimization model in this invention can be widely used in multi-robot collaborative path optimization tasks, especially in scenarios such as hotel monitoring, inspection, and unmanned driving. In these scenarios, multiple robots need to collaborate in shared resources (such as paths) while avoiding path conflicts and ensuring that high-risk areas are covered first.

[0060] Step S4 mainly focuses on using the behavior recognition results to adjust and optimize the inspection path in real time during the inspection process. In step S2, the inspection path is preliminarily determined through a multi-objective optimization algorithm. However, in the actual inspection process, due to the dynamic changes in the environment and the occurrence of abnormal behaviors, the system needs to dynamically adjust the original path to ensure that the abnormal behavior area can be covered in time.

[0061] The present invention realizes adaptive path optimization during the inspection process by combining abnormal behavior identification with a path adjustment mechanism. Specifically, the inspection path will be adjusted according to real-time feedback data using the node risk weights and behavior prediction models updated in step S3. In this way, the system can respond to environmental changes in a timely manner and ensure priority inspections in high-risk areas, thereby improving inspection efficiency and safety.

[0062] In step S4, the path adjustment mechanism is triggered in the following situations: Abnormal behavior occurs: When the inspection robot identifies abnormal behavior in a certain area (such as fighting, breaking windows, etc.), the risk weight of the area is R i It will be updated in step S3. Then, based on the updated risk data, the system will re-evaluate the inspection path.

[0063] Node risk weight change: When the system dynamically updates the node risk weight during the inspection process, the node v i The risk weight of Updated to The path planning system will then optimize the inspection path based on the new risk data.

[0064] It should be noted that path adjustment is a feedback mechanism based on real-time data, and the path planning algorithm will automatically adjust after each update to ensure the efficiency and safety of the path.

[0065] The path replanning algorithm used in this embodiment is based on the objective function f obtained in step S2. 1 (P), f 2 (P), f 3(P) is dynamically adjusted. When abnormal behavior occurs or risk weights are updated, the system will re-optimize the inspection path. The core objectives of path planning include: In the new path planning, the system will reduce the new path generation time as much as possible to minimize the inspection time.

[0066] The system will prioritize inspections of high-risk areas to ensure that abnormal areas are covered in a timely manner.

[0067] Avoid repeated coverage of areas that have already been inspected to ensure effective use of inspection resources.

[0068] In a possible implementation, the optimization of the inspection path can be dynamically balanced through the following optimization objectives: Among them, f 1 ′(P) represents the total inspection time of replanning path P, w i ' j represents the updated path edge weight, based on the new risk weight, f 2 ′(P) represents the total risk value of uncovered nodes, R i ′ is the updated node risk weight; By recalculating the time, risk coverage and redundancy of the path during the optimization process, the real-time effectiveness of the inspection path can be ensured.

[0069] During the path adjustment process, the system ensures that the inspection path is consistent with the real-time environment through a dynamic feedback mechanism. Specifically, when the risk of a node changes significantly (such as an event causing an increase in risk), the risk weight R of the node is increased. i It will be dynamically adjusted and fed back to the path planning module.

[0070] For example, when node v is identified during inspection i If abnormal behavior occurs (such as fire, fighting, etc.), the system will recalculate path P according to the new risk value, increase the inspection priority of the high-risk area, and ensure that the inspection of the area is completed in the shortest time.

[0071] in, is the updated node v i Risk weights, is the original node risk weight, ΔR i is the incremental risk caused by abnormal behavior, α and β are the smoothing coefficient and adjustment coefficient, which control the impact of historical data and new data.

[0072] The path planning module will regenerate the optimal inspection path P after each update and feed it back to the robot for execution.

[0073] This embodiment integrates deep learning and reinforcement learning technologies to make the path adjustment process more intelligent. During the inspection process, the system can not only adjust the path in real time, but also predict possible abnormal behaviors in the future by learning historical data and respond in advance. Through reinforcement learning, the system can continuously optimize the path planning strategy based on historical inspection feedback.

[0074] After each route update, the inspection robot will execute the new inspection route and feedback the inspection results and abnormal behaviors to the system. The system will further adjust the route based on this feedback data to ensure that subsequent inspection tasks can be more efficient and accurate.

[0075] Step S5 mainly involves feeding back the recognition results to the system during the abnormal behavior recognition process, and dynamically optimizing the entire system based on the results. The recognition of abnormal behavior is not only used for real-time response and adjustment of the inspection path, but also provides important data support for subsequent model training, risk assessment and inspection strategies. In the present invention, the system optimizes the planning and adjustment of the inspection path by feedback on the abnormal behavior recognition results, and improves the adaptability and efficiency of the entire inspection process.

[0076] The focus of this embodiment is to upload the identified abnormal behavior data to the cloud for further processing, and through data analysis and feedback mechanisms, dynamically adjust the behavior prediction model and node risk weights to ensure that the system can continue to improve as the inspection tasks proceed, thereby adapting to new challenges and environmental changes.

[0077] In step S5, when the inspection robot detects abnormal behavior through cameras and sensors, such as illegal intrusion, fighting, fire, etc., the robot will immediately upload the data to the cloud server. The uploaded data contains the time, location, type of behavior, and related multimodal information such as images and sounds of the abnormal behavior.

[0078] It should be noted that, in some embodiments, the uploaded data may also include changes in the surrounding environment (such as temperature changes, smoke concentration, etc.) and the robot's own status data. This information will help to further analyze the nature of the abnormal behavior and enhance the system's learning ability.

[0079] After the cloud server receives the abnormal behavior recognition results, the system first analyzes and processes the uploaded data. This analysis process includes: Data cleaning and preprocessing, noise removal, missing value supplementation and other steps; Through image recognition and sound analysis, we can further extract the specific features of abnormal behavior and identify behavioral patterns using deep learning algorithms; By analyzing the frequency and severity of abnormal behaviors and the environmental context in which the incident occurred, the potential risk impact of the incident on surrounding nodes is assessed to determine whether the risk weight needs to be adjusted.

[0080] If the system identifies a violent incident at the hotel front desk and the incident poses a significant threat to security, the system will reassess the risk weight of the front desk based on the severity of the incident and prioritize inspections of this area during route planning.

[0081] After completing the analysis and risk assessment of abnormal behavior, the system will dynamically adjust the risk weight of the node based on the analysis results. For example, for areas where serious abnormal behavior occurs, the risk weight of the node R i will be increased, thereby increasing the inspection priority of the area. The update process can be completed by the following formula: in, is the updated node v i Risk weights, is the original node v i Risk weight, γ is the incremental risk adjustment coefficient, which is used to control the magnitude of risk weight update, ΔR i is the incremental risk caused by abnormal behavior.

[0082] The abnormal behavior recognition results will also be fed back to the behavior prediction model for further training and optimization of model parameters. By regularly updating the training data set, the system can extract patterns from a large amount of historical inspection data and behavior recognition results, further improving the model's prediction capabilities.

[0083] In one possible implementation, a deep learning model such as a recurrent neural network (RNN) or a long short-term memory network (LSTM) is used to learn the behavior sequence to capture the temporal dependencies of the behavior. The optimization of the behavior prediction model enables the system to more accurately predict abnormal behaviors that may occur in the future and adjust the inspection strategy in advance. For example, if the system recognizes that a certain abnormal behavior (such as fighting) frequently occurs at a certain node, the model will predict the probability of the behavior and plan high-risk paths in advance based on the prediction results for priority inspection.

[0084] Based on data upload, analysis and processing, the system will dynamically adjust the model's prediction parameters to ensure that the behavior prediction model is consistent with the real-time needs of the inspection task. Specifically, the optimization process may involve the following aspects: dynamically adjusting the weights of the neural network through feedback from training data, updating hyperparameters (such as learning rate, loss function, etc.) to improve the generalization ability of the model, and adding newly identified abnormal behavior features to model training to make the model more comprehensive.

[0085] Once the risk weight of the node is updated or the behavior prediction model parameters are adjusted, the system will re-plan the inspection path based on these updated results. The updated inspection path will give priority to covering high-risk areas to ensure that areas where abnormal behavior occurs are inspected in a timely manner. In some embodiments, the system will also continuously evaluate and optimize the inspection strategy through reinforcement learning algorithms, so that the inspection path planning not only considers the current environment, but also predicts potential risks in the future, thereby achieving long-term optimization.

[0086] Please refer to the attached Figure 2 The present invention also provides an abnormal behavior identification system based on hotel monitoring and dynamic inspection, comprising: Inspection robot: used to perform inspection tasks, equipped with a variety of sensors, cameras and behavior recognition modules, to collect environmental data and video data, and analyze abnormal behavior in real time; Fixed surveillance cameras: deployed at multiple nodes within the hotel to collect static surveillance images and provide auxiliary inspection data; Path planning module: used to receive node risk weight information and generate the optimal inspection path of the inspection robot based on a multi-objective optimization algorithm; Collaboration optimization module: used to optimize the multi-robot collaboration strategy based on the game model and avoid inspection conflicts through dynamic path adjustment; Abnormal behavior identification module: used to identify abnormal behaviors of the collected inspection data based on the spatiotemporal feature extraction mechanism and the long short-term memory network model, and generate abnormal behavior classification results; Data processing and feedback module: used to receive inspection results and identification data, dynamically adjust node risk weights and train and update behavior prediction models to complete system parameter optimization; Cloud server: used to store inspection data, recognition results and historical behavior samples, and supports remote updating and deployment of path planning optimization and behavior prediction models.

[0087] Among them, the inspection robot is the execution unit in the system, responsible for dynamic inspection tasks in various areas of the hotel. Each inspection robot is equipped with advanced sensors and processing units, which can collect environmental data in real time and conduct preliminary analysis. The robot uses autonomous navigation technology to ensure that the task is executed according to the predetermined path, and responds to environmental changes in real time to ensure comprehensive and efficient monitoring; The fixed surveillance camera module provides static surveillance information for the hotel area. This module forms a static surveillance network covering all areas inside the hotel by deploying multiple surveillance cameras at key locations. It also cooperates with the patrol robot to enhance coverage and monitoring of the entire environment. The main function of the path planning module is to generate the optimal inspection path for the inspection robot based on the risk weight information of the node and the environmental status. This module uses a multi-objective optimization algorithm to dynamically adjust and optimize the path based on the risk level of each area, inspection efficiency, and feasibility of robot operation; The collaborative optimization module optimizes multi-robot collaborative tasks through a game optimization model to ensure that multiple robots can work together efficiently. This module avoids path conflicts and ensures priority coverage of high-risk areas by simulating interactions and resource competition between robots. The abnormal behavior recognition module is used to detect and identify abnormal behaviors based on the collected video and sensor data using a deep learning model. This module can accurately identify possible abnormal behaviors, such as fighting, stealing, smoking, etc., through spatiotemporal feature extraction and behavior prediction models; The data processing and feedback module is the core decision-making module of the system, which is responsible for receiving inspection results, behavior recognition data and output information of other modules, and dynamically adjusting system parameters based on these data. This module ensures that the system can perform adaptive optimization according to real-time changes; The cloud server is used to store all inspection data, recognition results, historical behavior samples and relevant models of the system, and supports remote updates and optimizations of the system. The cloud server is the data center of the entire system, ensuring long-term optimization and data management of the system.

[0088] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying abnormal behavior based on hotel monitoring and dynamic inspection, characterized in that: The following steps are involved: Model the hotel area, divide the monitoring range into multiple nodes, and calculate the risk weight of each node based on historical behavior data and real-time collected data; Based on the risk weight, the optimal inspection path of the inspection robot is generated through the path planning optimization algorithm; In multi-robot collaboration scenarios, the inspection path is dynamically adjusted based on the game optimization model to avoid inspection resource conflicts; The inspection robot collects images and environmental data in real time along the optimal path, and identifies abnormal behaviors based on spatiotemporal feature extraction and behavior prediction models; The identified abnormal behavior data is uploaded to the cloud, and the node risk weights and behavior prediction model parameters are dynamically updated.

2. The abnormal behavior identification method based on hotel monitoring and dynamic inspection according to claim 1 is characterized in that: The calculation of the risk weight of each node includes the following steps: Based on historical inspection data, the frequency of abnormal behavior of each node is counted to obtain the historical risk weight; Based on the real-time data collected by the inspection robot and fixed cameras, the risk value of the current node is predicted in combination with the abnormal behavior recognition model to obtain the real-time risk weight; The historical risk weights and real-time risk weights are integrated in a predetermined ratio to obtain a comprehensive risk weight.

3. The abnormal behavior identification method based on hotel monitoring and dynamic inspection according to claim 2 is characterized in that: The calculation formula for the comprehensive risk weight is: Among them, α and β are predetermined weight coefficients, which represent the weight ratio of historical data and real-time data in the comprehensive risk weight, and satisfy α+β=1. Represents the historical risk weight of the node, which is used to reflect the frequency of abnormal behavior of the node in the historical behavior data. Indicates the real-time risk weight of the node, which is used to reflect the probability of abnormal behavior of the node at the current moment.

4. The abnormal behavior identification method based on hotel monitoring and dynamic inspection according to claim 1 is characterized in that: The steps of the path planning optimization algorithm include: The monitoring nodes in the hotel environment are constructed into a graph structure, where the nodes represent the inspection areas and the edge weights represent the inspection time; Setting multi-objective optimization goals, including minimizing the total time of the inspection path, minimizing the risk value of uncovered high-risk nodes, and minimizing redundant coverage of the path; Use non-dominated sorting genetic algorithm to solve and generate Pareto solution set according to non-dominated sorting; The optimal solution that balances path time, risk coverage and redundancy is selected from the Pareto solution set as the inspection path.

5. The abnormal behavior identification method based on hotel monitoring and dynamic inspection according to claim 4 is characterized in that: The solution using the non-dominated sorting genetic algorithm comprises the following steps: Randomly generate several initial inspection paths and calculate the evaluation value of each path on the multi-objective function; Performance on all targets,classifies the population into several non-dominated classes; The crowding distance is calculated for each individual to evaluate population diversity; Select individuals with higher non-dominant ranks and larger crowding distances from the population as parents; Perform a crossover operation on the selected parent generation to generate offspring, and perform a mutation operation on the offspring; Merge the parent and child generations and select a new population based on non-dominated sorting and crowding distance; Repeat the above steps until the preset number of iterations and optimization conditions are met.

6. The abnormal behavior identification method based on hotel monitoring and dynamic inspection according to claim 1 is characterized in that: The method of dynamically adjusting the inspection path based on the game optimization model specifically includes the following steps: A non-cooperative game model of patrol robot collaboration is constructed, where each robot calculates its own path benefit based on a benefit function, which includes the total risk benefit of path coverage and the penalty cost of path conflict; The optimal hybrid strategy for each robot to select a path is solved by iteratively updating the probability distribution of each robot's path selection, and the probability distribution of the path selection is dynamically optimized by the benefit gradient method; After determining the path of each robot, the optimal path combination of all inspection robots under the game model is obtained by solving the Nash equilibrium, which avoids path conflicts and ensures priority coverage of high-risk areas.

7. The abnormal behavior identification method based on hotel monitoring and dynamic inspection according to claim 6 is characterized in that: The Nash equilibrium satisfies the following conditions: in, For robot i in strategy and other robot strategies The following income, Roboti in other strategies and other robot strategies The income under S i is the path selection set of robot i, S -i The set selected for all other robot paths.

8. The abnormal behavior identification method based on hotel monitoring and dynamic inspection according to claim 1 is characterized in that: The abnormal behavior identification based on spatiotemporal feature extraction and behavior prediction model includes the following steps: Extract the temporal features of the image data collected by the inspection robot and use the time dimension attention mechanism to capture the dynamic changes of abnormal behaviors; Extract the spatial features of image data and use the spatial dimension attention mechanism to determine the area where abnormal behavior occurs; The temporal and spatial features are processed jointly to obtain the comprehensive feature weights; The comprehensive feature weights are input into the long short-term memory network model to capture the long-term dependencies of the behavior sequence; Based on the recognition results output by the model, the behavior is classified to determine whether it is abnormal behavior.

9. The abnormal behavior identification method based on hotel monitoring and dynamic inspection according to claim 1 is characterized in that: The abnormal behavior identification result is used to dynamically update the node risk weight and behavior prediction model parameters, including the following steps: adjusting the real-time risk weight of the corresponding node based on the abnormal behavior identification result; Upload the behavior data and recognition results collected during the inspection process to the cloud to generate a new behavior sample set; Retrain the behavior prediction model with a new behavior sample set and update the model parameters; The updated risk weights and model parameters are sent to the path planning module and abnormal behavior identification module for optimization of the next round of inspection tasks.

10. An abnormal behavior identification system based on hotel monitoring and dynamic inspection, characterized in that: include: Inspection robot: used to perform inspection tasks, equipped with a variety of sensors, cameras and behavior recognition modules, to collect environmental data and video data, and analyze abnormal behavior in real time; Fixed surveillance cameras: deployed at multiple nodes within the hotel to collect static surveillance images and provide auxiliary inspection data; Path planning module: used to receive node risk weight information and generate the optimal inspection path of the inspection robot based on a multi-objective optimization algorithm; Collaboration optimization module: used to optimize multi-robot collaboration strategies based on game models and avoid inspection conflicts through dynamic path adjustment; Abnormal behavior identification module: used to identify abnormal behaviors of collected inspection data based on spatiotemporal feature extraction mechanism and long short-term memory network model, and generate abnormal behavior classification results; Data processing and feedback module: used to receive inspection results and identification data, dynamically adjust node risk weights and train and update behavior prediction models to complete system parameter optimization; Cloud server: used to store inspection data, recognition results and historical behavior samples, and supports remote updating and deployment of path planning optimization and behavior prediction models.

Images