Personnel safety detection method and device based on dynamic resource allocation

Abstract

The invention relates to a personnel security detection method and device based on dynamic resource allocation, which do not need additional hardware acceleration equipment, can realize multi-region high-precision monitoring by depending on a single host, and reduces the deployment cost. The method comprises the following steps: extracting video frames from camera video streams respectively corresponding to a plurality of security restricted areas; according to the first reasoning frame rate, frame reasoning tasks corresponding to the multiple safety limited areas are executed, and a personnel safety detection result is obtained; under the condition that the personnel safety detection result represents that the personnel exist in the safety limited area, determining an abnormal safety limited area in which the personnel exist; if yes, the execution of the frame reasoning tasks corresponding to other safety limited areas is suspended, the frame reasoning task corresponding to the abnormal safety limited area is executed according to a second reasoning frame rate, a new personnel safety detection result corresponding to the abnormal safety limited area is obtained, and the second reasoning frame rate is higher than the first reasoning frame rate.

Description

Technical Field

[0001] This disclosure relates to the field of safety detection technology, specifically to a personnel safety detection method and device based on dynamic resource allocation. Background Technology

[0002] In industrial production, warehousing and logistics, and public facilities, personnel safety monitoring is a crucial aspect of ensuring production order and personal safety. Traditional monitoring systems typically rely on manual monitoring or simple motion detection algorithms, but these methods suffer from low efficiency, high false alarm rates, and an inability to handle complex scenarios in real time. In recent years, neural network-based machine vision technology has been increasingly applied to personnel safety monitoring. By capturing video streams through cameras and utilizing deep learning models for target detection and behavior analysis, automated and intelligent safety monitoring can be achieved.

[0003] However, existing technologies still have significant limitations in multi-area monitoring scenarios. For example, when multiple security-restricted areas need to be monitored simultaneously, the computing resources of GPUs (Graphics Processing Units) limit the ability of a single host to perform real-time inference across all areas at a high frame rate. Taking a common 3070 GPU as an example, a single frame inference takes approximately 10 milliseconds. If 20 areas need to be monitored, the allocated frame rate for each area is only 5 frames per second. This low frame rate inference may result in slow response times to abnormal situations (such as unauthorized intrusion), missing critical intervention opportunities. Summary of the Invention

[0004] To overcome the problems existing in the related technologies, this disclosure provides a personnel safety detection method and apparatus based on dynamic resource allocation to solve the defects in the related technologies.

[0005] According to a first aspect of the present disclosure, a personnel safety detection method based on dynamic resource allocation is provided, comprising:

[0006] Extract video frames from the camera video streams corresponding to multiple security-restricted areas;

[0007] According to the first inference frame rate, the frame inference tasks corresponding to the security-restricted areas are executed to obtain the personnel safety detection results corresponding to each security-restricted area. The frame inference task corresponding to the security-restricted area is used to infer the video frames corresponding to the security-restricted area through a pre-trained neural network. The personnel safety detection results are used to characterize whether there are personnel in the security-restricted area.

[0008] If the personnel safety detection results indicate the presence of personnel in a safety-restricted area, then an abnormal safety-restricted area with personnel is identified.

[0009] The execution of frame inference tasks corresponding to other security-restricted areas (excluding the abnormal security-restricted area) in multiple security-restricted areas is suspended. The frame inference task corresponding to the abnormal security-restricted area is executed according to the second inference frame rate to obtain a new personnel security detection result corresponding to the abnormal security-restricted area. The second inference frame rate is higher than the first inference frame rate.

[0010] In some embodiments, performing the frame inference task corresponding to the abnormal security-restricted region according to the second inference frame rate includes:

[0011] In response to the detection of the abnormal security-restricted area, the abnormal security-restricted area is added to the abnormal queue, and it is determined whether the length of the abnormal queue exceeds the maximum number of abnormal areas threshold.

[0012] If the length of the anomaly queue does not exceed the maximum number of anomaly regions threshold, then the frame inference task corresponding to the anomaly security-restricted region is executed according to the second inference frame rate.

[0013] If the length of the anomaly queue exceeds the maximum number of anomaly regions threshold, then the inference priorities of multiple anomaly security-restricted regions are determined, and the anomaly security-restricted regions in the anomaly queue are removed in order of inference priority from low to high, until the length of the anomaly queue is equal to the maximum number of anomaly regions threshold. Then, the frame inference task corresponding to the anomaly security-restricted region in the anomaly queue is executed according to the second inference frame rate.

[0014] In some embodiments, the maximum number of abnormal regions threshold is obtained by rounding down the result of dividing the maximum inference frame rate of the graphics processor performing the inference task by the minimum frame rate required by the target tracking algorithm in the inference task.

[0015] In some embodiments, determining the inference priority of the plurality of anomalous security-restricted regions includes:

[0016] For each of the abnormal security restricted areas, the abnormal confidence level, abnormal category, and abnormal duration of the abnormal security restricted area output by the neural network are determined, and the abnormal confidence level, abnormal category, and abnormal duration are weighted and calculated based on a first preset weight to obtain the abnormality degree of the abnormal security restricted area, wherein the abnormal category is used to characterize the category of personnel behavior in the abnormal security restricted area;

[0017] For each of the aforementioned abnormal security-restricted areas, the regional asset value, personnel density, historical anomaly probability, and compliance level of the abnormal security-restricted area are determined. Based on a second preset weight, the regional asset value, personnel density, historical anomaly probability, and compliance level are weighted and calculated to obtain the business importance of the abnormal security-restricted area.

[0018] The reasoning priority of the multiple abnormal security-restricted regions is determined based on the degree of abnormality and the degree of business importance corresponding to each of the multiple abnormal security-restricted regions.

[0019] In some embodiments, it also includes:

[0020] The time decay constant and scene sensitivity coefficient corresponding to the multiple abnormal security-restricted areas are obtained respectively. The time decay constant is inversely proportional to the resource load of the image processor executing the inference task. The scene sensitivity coefficient is obtained through reinforcement learning. The reward value of the reinforcement learning is obtained by weighting the risk of equipment loss, personnel loss, and compliance risk.

[0021] The step of determining the inference priority of multiple abnormal security-restricted regions based at least on the degree of abnormality and the degree of business importance corresponding to each of the multiple abnormal security-restricted regions includes:

[0022] For each of the aforementioned abnormal security-restricted areas, the abnormal decay time value corresponding to the abnormal security-restricted area is determined based on the time decay constant and abnormal duration corresponding to the abnormal security-restricted area. The abnormality degree, business importance, scenario sensitivity coefficient and abnormal decay time value of the abnormal security-restricted area are multiplied together to obtain the priority score of the abnormal security-restricted area.

[0023] Based on the priority score, the inference priority of multiple abnormal security-restricted regions is determined, wherein the priority score is proportional to the inference priority.

[0024] In some embodiments, performing the frame inference task corresponding to the abnormal security-restricted region according to the second inference frame rate includes:

[0025] According to the second inference frame rate, extract the appearance and motion features of the target in the abnormal security-restricted area, and encode the target features using a depth feature descriptor;

[0026] Based on the Kalman filter and the feature encoding, the motion trajectory of the target is predicted, and based on the predicted motion trajectory, the position change and velocity change of the target in consecutive frames are calculated to establish a motion model;

[0027] Based on the motion model, abnormal targets that appear discontinuously or have abnormal motion trajectories are identified.

[0028] Filter out the abnormal targets.

[0029] In some embodiments, performing the frame inference task corresponding to the abnormal security-restricted region according to the second inference frame rate includes:

[0030] Based on the frequency of anomalies within a preset historical time period, the utilization rate of the graphics processor used to perform the inference task, and the memory bandwidth usage, the remaining resource usage time of the graphics processor is predicted.

[0031] If the remaining resource usage time is less than a preset duration threshold, the second inference frame rate is used as the initial target inference frame rate, and the following loop process is executed: the target inference frame rate is subtracted from the preset frame rate threshold to obtain the third inference frame rate. Based on the third inference frame rate, the frequency of anomalies, the utilization rate of the graphics processor, and the memory bandwidth usage, the new remaining resource usage time of the graphics processor is predicted. If the new remaining resource usage time is less than the preset duration threshold, the third inference frame rate is used as the new target inference frame rate until the new remaining resource usage time is greater than or equal to the preset duration threshold.

[0032] Execute the frame inference task corresponding to the abnormal security-restricted region according to the third inference frame rate at the end of the loop process.

[0033] In some embodiments, it also includes:

[0034] The inference tasks corresponding to the multiple security-restricted regions can be allocated to different threads using a multi-threaded framework, wherein one thread is used to execute frame inference tasks corresponding to one or more security-restricted regions; or, the inference tasks corresponding to the multiple security-restricted regions can be allocated to different streams using the CUDA streaming mechanism, wherein one stream is used to execute frame inference tasks corresponding to one or more security-restricted regions.

[0035] According to a second aspect of the present disclosure, a personnel safety detection device based on dynamic resource allocation is provided, comprising:

[0036] The frame extraction module is used to extract video frames from the camera video streams corresponding to multiple security-restricted areas.

[0037] The first detection module is used to execute frame inference tasks corresponding to multiple security-restricted areas according to a first inference frame rate, and obtain personnel safety detection results corresponding to each security-restricted area. The frame inference task corresponding to the security-restricted area is used to infer the video frames corresponding to the security-restricted area through a pre-trained neural network, and the personnel safety detection results are used to characterize whether there are personnel in the security-restricted area.

[0038] Anomaly determination module, used to determine an abnormal safety-restricted area where personnel are present when the personnel safety detection results indicate that personnel are present in the safety-restricted area;

[0039] The second detection module is used to pause the execution of frame inference tasks corresponding to other security-restricted areas in the multiple security-restricted areas except for the abnormal security-restricted area, and execute the frame inference task corresponding to the abnormal security-restricted area according to the second inference frame rate to obtain a new personnel security detection result corresponding to the abnormal security-restricted area, wherein the second inference frame rate is higher than the first inference frame rate.

[0040] According to a third aspect of the present disclosure, an electronic device is provided, the electronic device including a memory and a processor, the memory being configured to store computer instructions executable on the processor, and the processor being configured to implement the method described in any one of the first aspects when executing the computer instructions.

[0041] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0042] The personnel security detection method based on dynamic resource allocation provided in this disclosure dynamically adjusts the inference resource allocation according to the real-time status of a region by combining a first inference frame rate and a second inference frame rate. First, under normal monitoring conditions, all regions are inferred uniformly at a lower first inference frame rate. Once an anomaly is detected in a region (such as intrusion), the system immediately switches to the second inference frame rate, concentrating computing resources on that region. Thus, by dynamically adjusting the inference frame rate, resource waste caused by fixed frame rate allocation is avoided, enabling a single host to monitor more regions, improving GPU resource utilization, detection accuracy, and response speed. This method is suitable for large-scale deployment in scenarios such as industrial production, warehousing and logistics, and public facilities. Furthermore, this method requires no additional hardware acceleration equipment and can achieve high-precision monitoring of multiple regions using only a single host, reducing deployment costs. Attached Figure Description

[0043] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0044] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present disclosure of a personnel safety detection method based on dynamic resource allocation;

[0045] Figure 2 This is a system workflow diagram of a personnel safety detection method based on dynamic resource allocation, as illustrated in an exemplary embodiment of this disclosure.

[0046] Figure 3 This is a schematic diagram illustrating the dynamic inference frame rate adjustment process in a personnel safety detection method based on dynamic resource allocation, as shown in an exemplary embodiment of this disclosure.

[0047] Figure 4 This is a schematic diagram of the structure of a personnel safety detection device based on dynamic resource allocation, as shown in an exemplary embodiment of this disclosure;

[0048] Figure 5 This is a structural block diagram of an electronic device illustrated in an exemplary embodiment of the present disclosure. Detailed Implementation

[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0050] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0051] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0052] As mentioned in the background section, existing technologies still have significant limitations in multi-area monitoring scenarios. Low frame rate inference can lead to slow response times to anomalies (such as unauthorized intrusion), missing critical intervention opportunities. Some solutions attempt to improve frame rates by reducing model complexity or using hardware acceleration devices, but these methods often sacrifice detection accuracy and are costly and difficult to deploy on a large scale. For example, while lightweight model technologies (such as MOBILENET and EFFICIENTNET) can improve inference speed, detection accuracy in complex scenarios will significantly decrease. Hardware acceleration devices (such as FPGAs (Field Programmable Gate Arrays) and dedicated AI (Artificial Intelligence) chips) can improve performance, but deployment costs are high and they have high compatibility requirements with existing systems.

[0053] Therefore, how to achieve high-precision personnel safety detection in multiple areas under limited computing resources remains a technical challenge that urgently needs to be solved.

[0054] Based on this, in a first aspect, at least one embodiment of this disclosure provides a personnel safety detection method based on dynamic resource allocation, please refer to the appendix. Figure 1 The diagram illustrates the process of the method, including steps S101 to S104.

[0055] In step S101, video frames are extracted from the camera video streams corresponding to the multiple security-restricted areas.

[0056] In step S102, according to the first inference frame rate, frame inference tasks corresponding to multiple security-restricted areas are executed to obtain personnel safety detection results for each security-restricted area. Specifically, the frame inference task corresponding to the security-restricted area is used to infer the video frames corresponding to the security-restricted area using a pre-trained neural network, and the personnel safety detection results are used to characterize whether personnel are present in the security-restricted area.

[0057] In step S103, if the personnel safety detection results indicate that there are personnel in the safety-restricted area, an abnormal safety-restricted area where personnel are present is identified.

[0058] In step S104, the frame inference tasks corresponding to other security-restricted areas (excluding the abnormal security-restricted area) in multiple security-restricted areas are suspended, and the frame inference tasks corresponding to the abnormal security-restricted area are executed according to the second inference frame rate to obtain the new personnel security detection results corresponding to the abnormal security-restricted area. The second inference frame rate is higher than the first inference frame rate.

[0059] For example, multiple cameras are distributed across different security-restricted areas, capturing video streams in real time. Each security-restricted area corresponds to a frame inference task, thus obtaining the personnel security detection result for each area. If the personnel security detection result for any security-restricted area indicates the presence of personnel, then an anomaly is determined to exist in that area. Therefore, under normal monitoring conditions, uniform inference is performed on all areas at a low frame rate (i.e., the first inference frame rate); once an anomaly is detected in a certain area (such as personnel intrusion), the system immediately switches to full-time inference mode (i.e., using the second inference frame rate), concentrating computational resources on that area, increasing the inference frame rate to 100 frames per second, thereby supporting high-precision target tracking and anomaly detection.

[0060] In some embodiments, in response to detecting an abnormal security-restricted region, the abnormal security-restricted region is added to an abnormal queue, and it is determined whether the length of the abnormal queue exceeds a maximum abnormal region number threshold. If the length of the abnormal queue does not exceed the maximum abnormal region number threshold, the frame inference task corresponding to the abnormal security-restricted region is executed according to the second inference frame rate. If the length of the abnormal queue exceeds the maximum abnormal region number threshold, the inference priority of multiple abnormal security-restricted regions is determined, and the abnormal security-restricted regions in the abnormal queue are removed in order of inference priority from low to high until the length of the abnormal queue is equal to the maximum abnormal region number threshold. Then, the frame inference task corresponding to the abnormal security-restricted region in the abnormal queue is executed according to the second inference frame rate.

[0061] Therefore, an anomaly region priority management mechanism is introduced, setting a maximum region count threshold for full-time inference. When multiple regions experience anomalies simultaneously, resources are allocated according to preset priority rules (such as anomaly severity, region importance, etc.) to ensure high frame rate inference in critical regions while avoiding resource exhaustion.

[0062] In some embodiments, the maximum number of outlier regions threshold is obtained by rounding down the result of dividing the maximum inference frame rate of the graphics processor performing the inference task by the minimum frame rate required by the target tracking algorithm in the inference task. That is, the maximum number of outlier regions threshold can be calculated using the following formula:

[0063] (1);

[0064] in, The threshold for the maximum number of abnormal regions. The maximum inference frame rate for a graphics processor performing inference tasks. This represents the minimum frame rate required by the target tracking algorithm in the inference task.

[0065] In some embodiments, determining the inference priority of multiple abnormal security-restricted regions includes: for each abnormal security-restricted region, determining the abnormal confidence level, abnormal category, and abnormal duration of the abnormal security-restricted region output by the neural network, and performing a weighted calculation on the abnormal confidence level, abnormal category, and abnormal duration based on a first preset weight to obtain the abnormality degree of the abnormal security-restricted region; for each abnormal security-restricted region, determining the regional asset value, personnel density, historical abnormal probability, and compliance level of the abnormal security-restricted region, and performing a weighted calculation on the regional asset value, personnel density, historical abnormal probability, and compliance level based on a second preset weight to obtain the business importance of the abnormal security-restricted region; and determining the inference priority of multiple abnormal security-restricted regions based at least on the abnormality degree and business importance corresponding to each of the multiple abnormal security-restricted regions.

[0066] Among them, the anomaly category is used to characterize the type of personnel behavior in an abnormal security-restricted area, such as personnel breaking in, personnel lingering, and personnel fighting.

[0067] For example, the first preset weight includes preset weights corresponding to anomaly confidence level, anomaly category, and anomaly duration, respectively. The second preset weight includes preset weights corresponding to regional asset value, personnel density, historical anomaly probability, and compliance level, respectively.

[0068] For example, the degree of abnormality in an abnormally secure area can be calculated using the following formula:

[0069] (2);

[0070] in, The degree of abnormality in the abnormal security-restricted area, For abnormal confidence levels, As an anomaly category, For the duration of the abnormality, , and These are the preset weights corresponding to the anomaly confidence level, anomaly category, and anomaly duration, respectively.

[0071] For example, the business importance of an abnormally secure restricted area can be calculated using the following formula:

[0072] (3);

[0073] in, Based on the business importance of the abnormally secure restricted area, For regional asset value For personnel density, This represents the historical anomaly probability. For compliance level, , , and These are preset weights corresponding to regional asset value, population density, historical anomaly probability, and compliance level.

[0074] In some embodiments, the method further includes: obtaining time decay constants and scene sensitivity coefficients corresponding to multiple abnormal security-restricted regions, wherein the time decay constants are inversely proportional to the resource load of the image processor performing the inference task, and the scene sensitivity coefficients are obtained through online adjustment using reinforcement learning, with the reward value of the reinforcement learning being based on equipment loss risk, personnel loss risk, and compliance risk. Accordingly, the inference priority of multiple abnormal security-restricted regions is determined at least based on the degree of abnormality and business importance corresponding to each of the multiple abnormal security-restricted regions, including: for each abnormal security-restricted region, determining the abnormal decay time value of the abnormal security-restricted region based on the time decay constant and abnormal duration corresponding to the abnormal security-restricted region, and multiplying the degree of abnormality, business importance, scene sensitivity coefficient, and abnormal decay time value of the abnormal security-restricted region to obtain a priority score for the abnormal security-restricted region; and determining the inference priority of multiple abnormal security-restricted regions based on the priority scores, wherein the priority scores are directly proportional to the inference priority.

[0075] It should be understood that if priority scores do not decay over time, "newly entered anomalies" and "anomalies that have been delayed for 10 minutes" will be considered equally important, increasing the risk of delayed detection. Furthermore, the risks of anomalies on the same production line differ drastically between "full daytime operation" and "nighttime maintenance." Using scenario sensitivity coefficients, it is possible to distinguish between "unauthorized entry into the hazardous chemical storage facility" and "on-duty personnel inspection," reducing false alarms or missed alarms.

[0076] For example, the reward value for reinforcement learning can be obtained by weighting the risks of equipment loss, personnel loss, and compliance. Equipment loss risk can be represented by equipment downtime minutes, personnel loss risk by the actual number of casualties caused by anomalies, and compliance risk by penalty points for violations. The weights for the weighted calculation can be set according to actual needs; for example, a weight of 0.6 can be set for equipment loss risk, 0.2 for personnel loss risk, and 0.2 for compliance risk. This disclosure does not impose any limitations on this.

[0077] For example, the abnormal decay time value can be either a linear decay value or an exponential decay value. Taking exponential decay as an example, the abnormal decay time value can be obtained using the following formula:

[0078] (4);

[0079] in, This represents the abnormal decay time value. For the duration of the abnormality, is the time decay constant.

[0080] For example, the priority score of an abnormal security-restricted area can be obtained using the following formula:

[0081] (5);

[0082] in, Score based on priority. This represents the scene sensitivity coefficient.

[0083] In some embodiments, according to a second inference frame rate, a frame inference task corresponding to an abnormal security-restricted region is performed, including: extracting the appearance and motion features of a target in the abnormal security-restricted region according to the second inference frame rate, and encoding the target features using a deep feature descriptor; predicting the target's motion trajectory based on a Kalman filter and feature encoding, and calculating the target's position and velocity changes in consecutive frames based on the predicted motion trajectory to establish a motion model; identifying abnormal targets that appear discontinuously or have abnormal motion trajectories based on the motion model; and filtering abnormal targets.

[0084] It should be understood that existing low frame rate inference technologies cannot effectively handle "ghosting" phenomena in dynamic scenes (such as false detections caused by changes in lighting or object movement), and it is difficult to implement target tracking algorithms, leading to a decrease in the continuity and accuracy of personnel detection. In this embodiment of the present disclosure, at the second inference frame rate, a target tracking algorithm (such as SORT or DEEPSORT) can be combined to continuously detect personnel, and "ghosting" interference can be eliminated through trajectory analysis and feature matching. By utilizing the continuous frame data generated by high frame rate inference, a motion trajectory model of the target is established, and targets that appear discontinuously are eliminated, thereby effectively suppressing false detections in dynamic scenes.

[0085] In some embodiments, performing a frame inference task corresponding to an anomaly-restricted region according to a second inference frame rate includes: predicting the remaining resource usage time of the graphics processor based on the frequency of anomalies occurring within a preset historical time, the utilization rate of the graphics processor used to perform the inference task, and the memory bandwidth usage; if the remaining resource usage time is less than a preset time threshold, using the second inference frame rate as the initial target inference frame rate, and performing a loop process: subtracting the preset frame rate threshold from the target inference frame rate to obtain a third inference frame rate; predicting a new remaining resource usage time of the graphics processor based on the third inference frame rate, the frequency of anomalies occurring, the utilization rate of the graphics processor, and the memory bandwidth usage; and if the new remaining resource usage time is less than the preset time threshold, using the third inference frame rate as the new target inference frame rate, until the new remaining resource usage time is greater than or equal to the preset time threshold; and performing a frame inference task corresponding to an anomaly-restricted region according to the third inference frame rate at the end of the loop process.

[0086] For example, the rate of decrease can be preset, such as 5 frames per second.

[0087] Therefore, the risk of resource depletion can be detected in advance, and high frame rate inference can be guaranteed as much as possible when resources are sufficient.

[0088] In some embodiments, a multi-threaded framework can be used to allocate inference tasks corresponding to multiple security-restricted regions to different threads, where one thread is used to execute frame inference tasks corresponding to one or more security-restricted regions; or, using the CUDA streaming mechanism, inference tasks corresponding to multiple security-restricted regions can be allocated to different streams, where one stream is used to execute frame inference tasks corresponding to one or more security-restricted regions. Thus, by using multi-threaded task scheduling to allocate inference tasks to different GPU streams (STREAM), parallel execution of inference tasks is ensured while reducing the overhead of task switching.

[0089] The personnel safety detection method based on dynamic resource allocation provided in this disclosure will be described below through another embodiment.

[0090] For example, the system architecture of this method includes the following core modules:

[0091] Camera distribution: Multiple cameras are distributed in different security-restricted areas to capture video streams in real time.

[0092] Edge host: Receives video streams from cameras, and is responsible for inference calculations and alarm triggering.

[0093] GPU Resource Manager: Dynamically allocates inference tasks and controls the inference frame rate of each region.

[0094] Alarm Management Module: Responsible for triggering and recording alarms for abnormal situations.

[0095] like Figure 2 As shown, the following is the workflow of the system:

[0096] 1. Initialization Phase: The camera establishes a connection with the edge host and begins video streaming. The host loads a pre-trained neural network model (such as YOLOv8 or EfficientDet). Security-restricted ranges for each area are configured (defined by coordinate points). The GPU resource manager is initialized, and dynamic inference parameters are set (such as uniform inference frame rate, full-time inference frame rate, maximum number of outlier regions, etc.).

[0097] 2. Uniform Inference Phase: The edge host extracts frames from the video streams of each camera at fixed intervals. A pre-trained neural network is used to infer each frame, detecting the bounding box (box coordinates) of personnel. If the detected personnel's box coordinates are within a pre-defined restricted area, an alarm is triggered, and full-time inference mode is entered. The inference frame rate is dynamically allocated based on GPU performance and the number of monitored areas (e.g., when a 3070 GPU monitors 20 areas, 5 frames / second are allocated per area).

[0098] 3. Full-Time Inference Phase: When an anomaly is detected in a region, the GPU Resource Manager immediately increases the inference frame rate for that region to 100 frames per second, while pausing inference tasks in other regions. In full-time inference mode, continuous detection of personnel is performed using target tracking algorithms (such as DeepSORT), and "ghosting" phenomena are eliminated through trajectory analysis. If anomalies occur in multiple regions simultaneously, resources are allocated according to preset priority rules to ensure high frame rate inference in critical regions.

[0099] 4. Resource recovery mechanism: When an abnormal area returns to normal (e.g., personnel leave the restricted area), the GPU resource manager switches that area back to uniform inference mode and resumes monitoring of other areas.

[0100] like Figure 3 As shown, the dynamic inference frame rate adjustment process in this disclosure includes a uniform inference frame rate (i.e., the first inference frame rate mentioned above) and a full-time inference frame rate (i.e., the second inference frame rate mentioned above). The inference frame rate allocation follows the calculation formula:

[0101] (6);

[0102] in, This refers to the actual inference frame rate used. For uniform inference frame rate (e.g., 5 frames / second). The full-time inference frame rate (e.g., 100 frames per second).

[0103] In full-time inference mode, target tracking algorithms are used to continuously detect people, and "ghosting" is eliminated through the following steps:

[0104] 1. Feature Extraction: Extract the appearance features (such as color and shape) and motion features (such as speed and direction) of the target. Encode the target features using deep feature descriptors (such as ResNet-34 in DeepSORT).

[0105] 2. Trajectory Modeling: Predict and update the target's trajectory based on a Kalman filter. Calculate the target's position and velocity changes in consecutive frames to establish a motion model.

[0106] 3. Ghosting Removal: Targets that appear discontinuously or have abnormal motion trajectories are marked as "ghosts" and removed. By setting a minimum frame threshold for consecutive appearances (e.g., 5 frames), false detections within a short period are filtered out.

[0107] To improve GPU resource utilization, this disclosure employs asynchronous inference and resource scheduling optimization techniques:

[0108] 1. Multi-threaded task scheduling: Use a multi-threaded framework (such as OpenMP or C++11 thread library) to distribute inference tasks to different threads. Each thread is responsible for the inference tasks of one or more cameras, ensuring that tasks are executed in parallel.

[0109] 2. GPU Stream Management: Utilizing the CUDA stream mechanism, inference tasks are distributed across different streams to achieve parallel execution. Asynchronous memory copying and kernel startup reduce the overhead of task switching.

[0110] 3. Task Priority Queue: Inference tasks in abnormal areas have the highest priority to ensure timely processing of critical tasks. Inference tasks in normal areas are allocated resources using a round-robin method to ensure monitoring coverage of all areas.

[0111] To achieve high-precision detection with limited resources, this invention combines model lightweighting and knowledge distillation techniques:

[0112] 1. Lightweight Model: Use lightweight detection models (such as YOLOv8-N or EfficientDet-Lite) as the base model to reduce inference latency. Further compress the model size and improve inference efficiency through model pruning and quantization techniques.

[0113] 2. Knowledge Distillation: This technique transfers feature representations from large models (such as YOLOv8-L) to lightweight models, maintaining high accuracy while reducing computational complexity. It uses soft labeling and feature map distillation methods to ensure that the detection performance of the lightweight model is close to that of the large model.

[0114] To reduce inference latency, this invention employs the following optimization strategy:

[0115] 1. Batch Inference: During the uniform inference phase, frame data from multiple regions is organized into batches and fed into the GPU for inference in batches, reducing inference overhead. The batch size is dynamically adjusted to balance latency and throughput.

[0116] 2. Memory management optimization: Utilize shared memory and zero-copy techniques to reduce data transfer overhead. Pre-allocate memory space required for inference to avoid latency caused by dynamic memory allocation.

[0117] 3. Operator Fusion Optimization: Continuous operators (such as convolution and activation functions) in the neural network are fused into a single computational unit, reducing kernel call overhead. CUDA's Tensor Core technology is used to accelerate matrix operations and improve inference efficiency.

[0118] To improve GPU resource utilization, this invention employs the following strategy:

[0119] 1. Load Balancing: Dynamically adjust the allocation of inference tasks across regions to ensure balanced GPU load. During the processing of abnormal regions, idle resources are dynamically reclaimed for other tasks.

[0120] 2. Task Prefetching and Caching: Prefetch and cache the next frame's data to reduce task switching waiting time. Utilize the GPU's high-speed cache to store frequently accessed data (such as model weights) to improve access speed.

[0121] 3. Asynchronous data processing: While the inference task is being executed, the next frame of data is processed asynchronously (such as decoding and preprocessing) to improve the overall throughput.

[0122] In a production line monitoring scenario at a factory, a personnel safety detection system based on this disclosure was deployed: Number of cameras: 20 cameras, covering all security-restricted areas. GPU configuration: Each host is equipped with a 3070 GPU. Uniform inference frame rate: 5 frames / second. Full-time inference frame rate: 100 frames / second. Maximum abnormal region threshold: 5.

[0123] The system performs as follows: Under normal monitoring conditions, all areas perform uniform inference at a frame rate of 5 frames per second, with no resource bottlenecks. When an intruder is detected in a certain area, the system immediately switches to a full-time inference mode of 100 frames per second. The target tracking algorithm successfully eliminates the "ghosting" phenomenon, achieving a detection accuracy of 99.5%. When anomalies occur in multiple areas simultaneously, the system allocates resources according to priority rules to ensure high frame rate inference in critical areas, preventing resource exhaustion.

[0124] In addition, this disclosure supports the following extensions:

[0125] 1. Multi-host cluster deployment: Through distributed task scheduling, it supports multiple hosts to work together, further expanding the number of monitoring areas.

[0126] 2. Model upgrade: Supports online updates of the detection model without restarting the system.

[0127] 3. Functional expansion: It can integrate behavioral analysis (such as running and climbing), identity recognition and other functions to meet the needs of more scenarios.

[0128] Furthermore, this disclosure is compatible with the following hardware and software platforms: Hardware platform: Supports NVIDIA GPUs (such as 3070, 3090, A100) and mainstream CPU architectures (such as x86, ARM). Software platform: Supports Windows and Linux operating systems, and is compatible with inference engines such as ONNX Runtime and TensorRT.

[0129] Therefore, through innovative technologies such as dynamic inference frame rate adjustment, abnormal area priority management, target tracking, and ghosting elimination, high-precision personnel safety detection in multiple areas is achieved with limited GPU resources. Compared with existing technologies, this invention significantly improves resource utilization, detection accuracy, and response speed, making it suitable for large-scale deployment in scenarios such as industrial production, warehousing and logistics, and public facilities. Furthermore, through system performance optimization and scalability design, it can flexibly adapt to monitoring needs of different scales and complexities, providing an efficient and low-cost solution for the field of personnel safety monitoring.

[0130] According to a second aspect of the embodiments of this disclosure, a personnel safety detection device based on dynamic resource allocation is provided. Please refer to the appendix. Figure 4 The personnel safety detection device 400 based on dynamic resource allocation includes:

[0131] The frame extraction module 401 is used to extract video frames from the camera video streams corresponding to multiple security-restricted areas.

[0132] The first detection module 402 is used to execute frame inference tasks corresponding to multiple security-restricted areas according to the first inference frame rate, and obtain personnel safety detection results corresponding to each security-restricted area. The frame inference task corresponding to the security-restricted area is used to infer the video frames corresponding to the security-restricted area through a pre-trained neural network, and the personnel safety detection results are used to characterize whether there are personnel in the security-restricted area.

[0133] Anomaly determination module 403 is used to determine an abnormal safety-restricted area where personnel exist when the personnel safety detection result indicates that personnel exist in the safety-restricted area.

[0134] The second detection module 404 is used to pause the execution of frame inference tasks corresponding to other security-restricted areas in the multiple security-restricted areas except for the abnormal security-restricted area, and execute the frame inference task corresponding to the abnormal security-restricted area according to the second inference frame rate to obtain a new personnel security detection result corresponding to the abnormal security-restricted area, wherein the second inference frame rate is higher than the first inference frame rate.

[0135] In some embodiments of this disclosure, the abnormal security restricted area is multiple, and the second detection module 404 is further used for:

[0136] In response to the detection of the abnormal security-restricted area, the abnormal security-restricted area is added to the abnormal queue, and it is determined whether the length of the abnormal queue exceeds the maximum number of abnormal areas threshold.

[0137] If the length of the anomaly queue does not exceed the maximum number of anomaly regions threshold, then the frame inference task corresponding to the anomaly security-restricted region is executed according to the second inference frame rate.

[0138] If the length of the anomaly queue exceeds the maximum number of anomaly regions threshold, then the inference priorities of multiple anomaly security-restricted regions are determined, and the anomaly security-restricted regions in the anomaly queue are removed in order of inference priority from low to high, until the length of the anomaly queue is equal to the maximum number of anomaly regions threshold. Then, the frame inference task corresponding to the anomaly security-restricted region in the anomaly queue is executed according to the second inference frame rate.

[0139] In some embodiments of this disclosure, the maximum number of abnormal regions threshold is obtained by rounding down the result of dividing the maximum inference frame rate of the graphics processor performing the inference task by the minimum frame rate required by the target tracking algorithm in the inference task.

[0140] In some embodiments of this disclosure, the second detection module 404 is further configured to:

[0141] For each of the abnormal security restricted areas, the abnormal confidence level, abnormal category, and abnormal duration of the abnormal security restricted area output by the neural network are determined, and the abnormal confidence level, abnormal category, and abnormal duration are weighted and calculated based on a first preset weight to obtain the abnormality degree of the abnormal security restricted area, wherein the abnormal category is used to characterize the category of personnel behavior in the abnormal security restricted area;

[0142] For each of the aforementioned abnormal security-restricted areas, the regional asset value, personnel density, historical anomaly probability, and compliance level of the abnormal security-restricted area are determined. Based on a second preset weight, the regional asset value, personnel density, historical anomaly probability, and compliance level are weighted and calculated to obtain the business importance of the abnormal security-restricted area.

[0143] The reasoning priority of the multiple abnormal security-restricted regions is determined based on the degree of abnormality and the degree of business importance corresponding to each of the multiple abnormal security-restricted regions.

[0144] In some embodiments of this disclosure, an acquisition module is also included, for:

[0145] The time decay constant and scene sensitivity coefficient corresponding to the multiple abnormal security-restricted areas are obtained respectively. The time decay constant is inversely proportional to the resource load of the image processor executing the inference task. The scene sensitivity coefficient is obtained through reinforcement learning. The reward value of the reinforcement learning is obtained by weighting the risk of equipment loss, personnel loss, and compliance risk.

[0146] The second detection module 404 is further used for:

[0147] For each of the aforementioned abnormal security-restricted areas, the abnormal decay time value corresponding to the abnormal security-restricted area is determined based on the time decay constant and abnormal duration corresponding to the abnormal security-restricted area. The abnormality degree, business importance, scenario sensitivity coefficient and abnormal decay time value of the abnormal security-restricted area are multiplied together to obtain the priority score of the abnormal security-restricted area.

[0148] Based on the priority score, the inference priority of multiple abnormal security-restricted regions is determined, wherein the priority score is proportional to the inference priority.

[0149] In some embodiments of this disclosure, the second detection module 404 is further configured to:

[0150] According to the second inference frame rate, extract the appearance and motion features of the target in the abnormal security-restricted area, and encode the target features using a depth feature descriptor;

[0151] Based on the Kalman filter and the feature encoding, the motion trajectory of the target is predicted, and based on the predicted motion trajectory, the position change and velocity change of the target in consecutive frames are calculated to establish a motion model;

[0152] Based on the motion model, abnormal targets that appear discontinuously or have abnormal motion trajectories are identified.

[0153] Filter out the abnormal targets.

[0154] In some embodiments of this disclosure, the second detection module 404 is further configured to:

[0155] Based on the frequency of anomalies within a preset historical time period, the utilization rate of the graphics processor used to perform the inference task, and the memory bandwidth usage, the remaining resource usage time of the graphics processor is predicted.

[0156] If the remaining resource usage time is less than a preset duration threshold, the second inference frame rate is used as the initial target inference frame rate, and the following loop process is executed: the target inference frame rate is subtracted from the preset frame rate threshold to obtain the third inference frame rate. Based on the third inference frame rate, the frequency of anomalies, the utilization rate of the graphics processor, and the memory bandwidth usage, the new remaining resource usage time of the graphics processor is predicted. If the new remaining resource usage time is less than the preset duration threshold, the third inference frame rate is used as the new target inference frame rate until the new remaining resource usage time is greater than or equal to the preset duration threshold.

[0157] Execute the frame inference task corresponding to the abnormal security-restricted region according to the third inference frame rate at the end of the loop process.

[0158] In some embodiments of this disclosure, a parallel module is also included for:

[0159] The inference tasks corresponding to the multiple security-restricted regions can be allocated to different threads using a multi-threaded framework, wherein one thread is used to execute frame inference tasks corresponding to one or more security-restricted regions; or, the inference tasks corresponding to the multiple security-restricted regions can be allocated to different streams using the CUDA streaming mechanism, wherein one stream is used to execute frame inference tasks corresponding to one or more security-restricted regions.

[0160] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments of the method in the first aspect, and will not be elaborated upon here.

[0161] According to a third aspect of the embodiments of this disclosure, please refer to the appendix. Figure 5 The diagram illustrates an exemplary block diagram of an electronic device 700, which may include a processor 701 and a memory 702. The electronic device 700 may also include one or more of a multimedia component 703, an input / output (I / O) interface 704, and a communication component 705.

[0162] The processor 701 controls the overall operation of the electronic device 700 to complete all or part of the steps in any of the methods described above. The memory 702 stores various types of data to support the operation of the electronic device 700. This data may include, for example, instructions for any application or method operating on the electronic device 700, and application-related data such as contact data, sent and received messages, pictures, audio, video, etc. The memory 702 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 703 may include a screen and audio components. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in memory 702 or transmitted via communication component 705. The audio component also includes at least one speaker for outputting audio signals. I / O interface 704 provides an interface between processor 701 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons may be virtual or physical buttons. Communication component 705 is used for wired or wireless communication between the electronic device 700 and other devices. Wireless communication may include Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof; therefore, the corresponding communication component 705 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0163] In an exemplary embodiment, the electronic device 700 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the aforementioned personnel safety detection method based on dynamic resource allocation.

[0164] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of any of the methods described above. For example, the computer-readable storage medium may be the memory 702 including the program instructions described above, which may be executed by the processor 701 of the electronic device 700 to perform any of the methods described above.

[0165] In another exemplary embodiment, a computer program product is also provided, which includes a computer program executable by a processor, wherein the computer program, when executed by the processor, implements the steps of any of the methods described above.

[0166] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0167] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0168] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A personnel safety detection method based on dynamic resource allocation, characterized in that, include: Extract video frames from the camera video streams corresponding to multiple security-restricted areas; According to the first inference frame rate, the frame inference tasks corresponding to the security-restricted areas are executed to obtain the personnel safety detection results corresponding to each security-restricted area. The frame inference task corresponding to the security-restricted area is used to infer the video frames corresponding to the security-restricted area through a pre-trained neural network. The personnel safety detection results are used to characterize whether there are personnel in the security-restricted area. If the personnel safety detection results indicate the presence of personnel in a safety-restricted area, then an abnormal safety-restricted area with personnel is identified. The execution of frame inference tasks corresponding to other security-restricted areas (excluding the abnormal security-restricted area) in multiple security-restricted areas is suspended. The frame inference task corresponding to the abnormal security-restricted area is executed according to the second inference frame rate to obtain a new personnel security detection result corresponding to the abnormal security-restricted area. The second inference frame rate is higher than the first inference frame rate.

2. The personnel safety detection method based on dynamic resource allocation according to claim 1, characterized in that, The step of performing the frame inference task corresponding to the abnormal security-restricted region according to the second inference frame rate includes: In response to the detection of the abnormal security-restricted area, the abnormal security-restricted area is added to the abnormal queue, and it is determined whether the length of the abnormal queue exceeds the maximum number of abnormal areas threshold. If the length of the anomaly queue does not exceed the maximum number of anomaly regions threshold, then the frame inference task corresponding to the anomaly security-restricted region is executed according to the second inference frame rate. If the length of the anomaly queue exceeds the maximum number of anomaly regions threshold, then the inference priorities of multiple anomaly security-restricted regions are determined, and the anomaly security-restricted regions in the anomaly queue are removed in order of inference priority from low to high, until the length of the anomaly queue is equal to the maximum number of anomaly regions threshold. Then, the frame inference task corresponding to the anomaly security-restricted region in the anomaly queue is executed according to the second inference frame rate.

3. The personnel safety detection method based on dynamic resource allocation according to claim 2, characterized in that, The maximum number of abnormal regions threshold is obtained by rounding down the result of dividing the maximum inference frame rate of the graphics processor executing the inference task by the minimum frame rate required by the target tracking algorithm in the inference task.

4. The personnel safety detection method based on dynamic resource allocation according to claim 2, characterized in that, The inference priority for determining the multiple abnormal security-restricted regions includes: For each of the abnormal security restricted areas, the abnormal confidence level, abnormal category, and abnormal duration of the abnormal security restricted area output by the neural network are determined, and the abnormal confidence level, abnormal category, and abnormal duration are weighted and calculated based on a first preset weight to obtain the abnormality degree of the abnormal security restricted area, wherein the abnormal category is used to characterize the category of personnel behavior in the abnormal security restricted area; For each of the aforementioned abnormal security-restricted areas, the regional asset value, personnel density, historical anomaly probability, and compliance level of the abnormal security-restricted area are determined. Based on a second preset weight, the regional asset value, personnel density, historical anomaly probability, and compliance level are weighted and calculated to obtain the business importance of the abnormal security-restricted area. The reasoning priority of the multiple abnormal security-restricted regions is determined based on the degree of abnormality and the degree of business importance corresponding to each of the multiple abnormal security-restricted regions.

5. The personnel safety detection method based on dynamic resource allocation according to claim 4, characterized in that, Also includes: The time decay constant and scene sensitivity coefficient corresponding to the multiple abnormal security-restricted areas are obtained respectively. The time decay constant is inversely proportional to the resource load of the image processor executing the inference task. The scene sensitivity coefficient is obtained through reinforcement learning. The reward value of the reinforcement learning is obtained by weighting the risk of equipment loss, personnel loss, and compliance risk. The step of determining the inference priority of multiple abnormal security-restricted regions based at least on the degree of abnormality and the degree of business importance corresponding to each of the multiple abnormal security-restricted regions includes: For each of the aforementioned abnormal security-restricted areas, the abnormal decay time value corresponding to the abnormal security-restricted area is determined based on the time decay constant and abnormal duration corresponding to the abnormal security-restricted area. The abnormality degree, business importance, scenario sensitivity coefficient and abnormal decay time value of the abnormal security-restricted area are multiplied together to obtain the priority score of the abnormal security-restricted area. Based on the priority score, the inference priority of multiple abnormal security-restricted regions is determined, wherein the priority score is proportional to the inference priority.

6. The personnel safety detection method based on dynamic resource allocation according to any one of claims 1-5, characterized in that, The step of performing the frame inference task corresponding to the abnormal security-restricted region according to the second inference frame rate includes: According to the second inference frame rate, extract the appearance and motion features of the target in the abnormal security-restricted area, and encode the target features using a depth feature descriptor; Based on the Kalman filter and the feature encoding, the motion trajectory of the target is predicted, and based on the predicted motion trajectory, the position change and velocity change of the target in consecutive frames are calculated to establish a motion model; Based on the motion model, abnormal targets that appear discontinuously or have abnormal motion trajectories are identified. Filter out the abnormal targets.

7. The personnel safety detection method based on dynamic resource allocation according to any one of claims 1-5, characterized in that, The step of performing the frame inference task corresponding to the abnormal security-restricted region according to the second inference frame rate includes: Based on the frequency of anomalies within a preset historical time period, the utilization rate of the graphics processor used to perform the inference task, and the memory bandwidth usage, the remaining resource usage time of the graphics processor is predicted. If the remaining resource usage time is less than a preset duration threshold, the second inference frame rate is used as the initial target inference frame rate, and the following loop process is executed: the target inference frame rate is subtracted from the preset frame rate threshold to obtain the third inference frame rate. Based on the third inference frame rate, the frequency of anomalies, the utilization rate of the graphics processor, and the memory bandwidth usage, the new remaining resource usage time of the graphics processor is predicted. If the new remaining resource usage time is less than the preset duration threshold, the third inference frame rate is used as the new target inference frame rate until the new remaining resource usage time is greater than or equal to the preset duration threshold. Execute the frame inference task corresponding to the abnormal security-restricted region according to the third inference frame rate at the end of the loop process.

8. The personnel safety detection method based on dynamic resource allocation according to any one of claims 1-5, characterized in that, Also includes: The inference tasks corresponding to the multiple security-restricted regions can be allocated to different threads using a multi-threaded framework, wherein one thread is used to execute frame inference tasks corresponding to one or more security-restricted regions; or, the inference tasks corresponding to the multiple security-restricted regions can be allocated to different streams using the CUDA streaming mechanism, wherein one stream is used to execute frame inference tasks corresponding to one or more security-restricted regions.

9. A personnel safety detection device based on dynamic resource allocation, characterized in that, include: The frame extraction module is used to extract video frames from the camera video streams corresponding to multiple security-restricted areas. The first detection module is used to execute frame inference tasks corresponding to multiple security-restricted areas according to a first inference frame rate, and obtain personnel safety detection results corresponding to each security-restricted area. The frame inference task corresponding to the security-restricted area is used to infer the video frames corresponding to the security-restricted area through a pre-trained neural network, and the personnel safety detection results are used to characterize whether there are personnel in the security-restricted area. Anomaly determination module, used to determine an abnormal safety-restricted area where personnel are present when the personnel safety detection results indicate that personnel are present in the safety-restricted area; The second detection module is used to pause the execution of frame inference tasks corresponding to other security-restricted areas in the multiple security-restricted areas except for the abnormal security-restricted area, and execute the frame inference task corresponding to the abnormal security-restricted area according to the second inference frame rate to obtain a new personnel security detection result corresponding to the abnormal security-restricted area, wherein the second inference frame rate is higher than the first inference frame rate.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory being used to store computer instructions executable on the processor, and the processor being used to implement the steps of the method according to any one of claims 1-8 when executing the computer instructions.

Images