Intelligent alarm system based on platform real-time video stream

Abstract

The invention discloses an intelligent alarm system based on a real-time video stream of a platform, and relates to the technical field of video processing early warning, and the intelligent alarm system comprises a data collection module which carries out the real-time monitoring based on a monitoring camera, reads the image data of a point position camera, carries out the standardization and normalization of the data, and carries out the unified input of the image data; and the target discovery module is used for inputting the preprocessed image frame and the monitored area, identifying and positioning pedestrians in the image frame by adopting an algorithm based on target detection, performing label processing on the identified pedestrians based on a detected tracking algorithm byterack, and outputting the residence time of the current frame target in the monitored area. According to the platform safety management system, various advantages are brought to platform safety management, the alarm threshold value can be flexibly adjusted, decision support is provided, management personnel are helped to analyze passenger flow, optimize deployment and troubleshoot hidden dangers by recording related data, a'early warning-processing-optimization 'closed loop is formed, the safety management and control level is improved, and safety accidents are effectively prevented.

Description

Technical Field

[0001] This invention relates to the field of video processing and early warning technology, specifically to an intelligent alarm system based on real-time video streams from a platform. Background Technology

[0002] In logistics scenarios, platforms are where trucks park and load / unload goods, facilitating efficient operation of the logistics system and rapid sorting and distribution of goods, making logistics centers more flexible. However, in recent years, due to management negligence, pedestrian accidents are prone to occur at platforms. Traditional monitoring methods rely on segmented frame-by-frame analysis of video streams for early warning, which means real-time alarms are not possible. Data processing and analysis based on historical frame-by-frame data results in delays in feedback to the center, leading to inaccurate alarms. The intermittent frame format also prevents filtering of pedestrians passing through the platform within a short period. The lack of a corresponding alarm mechanism results in delayed alarm feedback when problems occur. To address this, we propose an intelligent alarm system based on real-time video streams from platforms. Summary of the Invention

[0003] To address the aforementioned technical problems, an intelligent alarm system based on real-time video streams from the platform is provided. This technical solution resolves the problems described above.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent alarm system based on real-time video stream of a platform, the intelligent alarm system comprising: The data acquisition module, based on real-time monitoring by surveillance cameras, reads image data from the cameras at various locations, performs standardization and normalization on the data, and then inputs the image data in a unified manner. The target detection module takes a preprocessed image frame and the monitored area as input, uses a target detection-based algorithm to identify and locate pedestrians in the image frame, and uses the detection-based tracking algorithm bytetrack to label the identified pedestrians, outputting the dwell time of the target in the current frame within the monitored area. The alarm logic module takes the target dwell time in the current frame as input and judges the dwell time of different targets through multi-target analysis. If the dwell time exceeds the set threshold, it outputs a signal to control the alarm light to flash. When the target leaves the designated area, it outputs a signal to control the alarm light to stop working.

[0005] Preferably, the data standardization step involves converting color images to grayscale images to reduce data dimensionality and unify the color space; calculating the pixel mean of the image dataset; subtracting the mean from each pixel value to shift the data distribution center to 0; calculating the pixel standard deviation of the image dataset; dividing the centered data by the standard deviation to make the standard deviation of the data equal to 1, thus completing the standardization process; and normalization involves adjusting the size, converting the color format, and converting it to floating-point type; and scaling the pixel values ​​from 0-255 to the range [0, 1].

[0006] Preferably, preprocessing employs a distributed multi-threaded processing approach for the video stream, dynamically adjusting the processing frame rate based on camera network congestion using a distributed ABR algorithm. The specific steps of the ABR algorithm are as follows: Network status monitoring is performed. Edge nodes collect local bandwidth utilization, packet loss rate and round-trip latency data in real time. The central node statistically analyzes the global load balancing status and network topology. Based on the monitoring data, the network congestion level is graded and evaluated, and a congestion score is obtained through weighted calculation. Based on the congestion score, the video processing frame rate is dynamically adjusted, and three congestion levels—low, medium, and high—are defined, each corresponding to a different recommended frame rate range and processing strategy. The central node and edge nodes work together to perform global optimization. Edge nodes adjust the frame rate based on the local network status and report the status. The central node collects information from all nodes, builds a global network view, and executes load balancing strategies when there is local congestion. When a sudden congestion is detected, the frame rate is reduced and a caching mechanism is enabled; once network conditions improve, the frame rate is gradually increased and processing performance is verified. The algorithm execution cycle includes collecting network status data every 200-500ms, making frame rate adjustment decisions every 1-2s, and performing distributed collaborative optimization every 5-10s.

[0007] Preferably, the target detection algorithm is YOLO real-time target detection, specifically as follows: Load the pre-trained YOLO model and divide the entire image frame into multiple grids of equal size, with each grid responsible for predicting and identifying pedestrians; Based on each grid, multiple bounding boxes are predicted. The bounding boxes are used to enclose existing pedestrians. Each bounding box corresponds to a confidence score, which reflects both the probability of a pedestrian being inside the bounding box and the degree of fit between the bounding box and the actual location of the pedestrian. Each grid cell predicts the probability that the target within that grid cell belongs to the pedestrian category. During the prediction process, the category probability is multiplied by the confidence score of the bounding box to obtain the final score of each bounding box. The final score takes into account both the probability of the target's existence and the accuracy of the category judgment. The YOLO algorithm uses non-maximum suppression (NMS) to select high-scoring bounding boxes, suppress low-scoring bounding boxes that have a high degree of overlap with high-scoring bounding boxes, and retain the bounding boxes that accurately capture pedestrians. It identifies all pedestrians in the image frame and locates the position of each pedestrian in the image based on the finally determined bounding boxes.

[0008] Preferably, the detection-based tracking algorithm bytetrack performs label processing on the identified pedestrians and outputs the dwell time of the target in the current frame within the monitored area in the following steps: Based on previously obtained pedestrian information, bytetrack initiates the tracking process. When new image frames are received, it filters the detected pedestrian targets and distinguishes between high-confidence and low-confidence detection results. For high-confidence targets, the algorithm matches the high-confidence targets with existing tracking trajectories, calculates the similarity of appearance features and the predicted value of the motion trajectory, determines whether they are the same pedestrian, and updates the label information of the trajectory if the match is successful. For low-confidence targets, the low-confidence targets are used as candidate targets and are matched with temporarily lost trajectories. When a brand-new pedestrian target is detected and cannot be matched with any existing trajectory, a unique label is assigned to it and a new tracking trajectory is established.

[0009] Preferably, for each tagged pedestrian trajectory, bytetrack monitors its position in the image frame in real time. When a pedestrian first enters the monitored area, the algorithm starts timing and records the entry time. In subsequent consecutive frames, the pedestrian remains within the monitored area, and the timer continues to accumulate. If the pedestrian leaves the monitored area, the timer pauses. The accumulated time at this point is the time the pedestrian spent within the monitored area. Upon reaching the current frame, the algorithm, based on real-time location determination, outputs the corresponding pedestrian label and the cumulative dwell time within the monitored area, thus tracking and recording the pedestrian's dwelling status.

[0010] Preferably, the operation steps within the alarm logic module are as follows: It continuously receives real-time dwell time data of each pedestrian target in the current frame and associates it with the corresponding target label; A preset dwell time threshold is matched for each target. By comparing the cumulative dwell time of each target with the corresponding threshold frame by frame, it is determined whether the alarm condition is triggered. If the dwell time of a target exceeds the threshold for the first time, a control signal is generated to drive the alarm light to flash as a warning. For targets that have triggered alarms, continuously monitor the target status. If the dwell time continues to increase, the alarm signal remains unchanged. An alarm is triggered simultaneously when multiple targets exceed the threshold at the same time.

[0011] Preferably, the further operation steps within the alarm logic module are as follows: When the target leaves the monitoring area, the position judgment function identifies the change in status. After confirming that the target has left, a stop signal is generated to stop the flashing of the alarm light. There are multiple alarm targets at the same time. Each target is handled independently. When a target leaves, its associated alarm signal is stopped, while other targets that are still in the area and exceed the threshold continue to keep the alarm.

[0012] Preferably, the alarm logic module ignores single state transitions for brief misjudgments of location caused by obstruction or rapid movement; and the dwell time is re-accumulated when the target re-enters the area after it has completely left.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention brings multiple advantages to platform safety management: full-process automation, 24-hour uninterrupted data collection, a target detection module that can track the status of multiple pedestrians in parallel, precise alarm logic to control alarm activation and deactivation, saving manpower and shortening response time, strong system adaptability to different equipment and complex environments, flexible adjustment of alarm thresholds to provide decision support, and the recording of relevant data to help managers analyze passenger flow, optimize deployment, and identify potential hazards, forming a closed loop of "early warning-processing-optimization" to improve safety management and effectively prevent safety accidents. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the workflow of the alarm system of the present invention. Detailed Implementation

[0015] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0016] Reference Figure 1 As shown, an intelligent alarm system based on real-time video stream from a platform is described. The intelligent alarm system includes: The data acquisition module, based on real-time monitoring by surveillance cameras, reads image data from the cameras at various locations, performs standardization and normalization on the data, and then inputs the image data in a unified manner. The target detection module takes a preprocessed image frame and the monitored area as input, uses a target detection-based algorithm to identify and locate pedestrians in the image frame, and uses the detection-based tracking algorithm bytetrack to label the identified pedestrians, outputting the dwell time of the target in the current frame within the monitored area. The alarm logic module takes the target dwell time in the current frame as input and judges the dwell time of different targets through multi-target analysis. If the dwell time exceeds the set threshold, it outputs a signal to control the alarm light to flash. When the target leaves the designated area, it outputs a signal to control the alarm light to stop working.

[0017] This application brings significant advantages to platform safety management through the coordinated operation of three major modules. The data acquisition module standardizes the processing of image data from different cameras, ensuring consistent input, improving the stability of subsequent analysis, and reducing hardware adaptation costs. The target detection module combines detection and tracking algorithms to accurately identify and locate pedestrians and output dwell time in real time, enabling parallel monitoring of multiple targets and overcoming the efficiency and accuracy limitations of manual monitoring. The alarm logic module dynamically judges dwell time, accurately triggering or terminating alarms, reducing invalid interference, and ensuring rapid response to abnormal situations. The entire process is fully automated, operating 24 hours a day, saving labor costs. Furthermore, data accumulation can assist in passenger flow analysis and hazard investigation, forming a management closed loop and comprehensively improving the level of platform safety control.

[0018] The data standardization steps involve converting color images to grayscale images to reduce data dimensionality and unify the color space; calculating the pixel mean of the image dataset and subtracting this mean from each pixel value to shift the data distribution center to 0; calculating the pixel standard deviation of the image dataset and dividing the centered data by this standard deviation to make the standard deviation of the data equal to 1, thus completing the standardization process; the normalization process involves adjusting the size, converting the color format and converting it to floating-point type; and scaling the pixel values ​​from 0-255 to the range [0, 1].

[0019] In the standardization process of this application, converting to grayscale reduces dimensionality, unifies the color space, and reduces computational complexity; subtracting the mean from the pixels makes the data center zero, and dividing by the standard deviation makes the standard deviation 1, which can eliminate interference from factors such as lighting, make the data distribution more stable, and improve the model convergence speed and detection accuracy. Normalization adjusts the size and color format, converts to floating-point, and shrinks the pixel values ​​to the [0, 1] range. This not only unifies the data scale and avoids the impact of different ranges of pixel values ​​on model training, but also facilitates efficient processing by subsequent algorithms, enhances the system's adaptability to image data in different devices and environments, and provides a more reliable input basis for target detection and tracking.

[0020] Preprocessing employs a distributed, multi-threaded approach to the video stream, dynamically adjusting the processing frame rate based on camera network congestion using a distributed ABR algorithm. The specific steps of the ABR algorithm are as follows: Network status monitoring is performed. Edge nodes collect local bandwidth utilization, packet loss rate and round-trip latency data in real time. The central node statistically analyzes the global load balancing status and network topology. Based on the monitoring data, the network congestion level is graded and evaluated, and a congestion score is obtained through weighted calculation. Based on the congestion score, the video processing frame rate is dynamically adjusted, and three congestion levels—low, medium, and high—are defined, each corresponding to a different recommended frame rate range and processing strategy. The central node and edge nodes work together to perform global optimization. Edge nodes adjust the frame rate based on the local network status and report the status. The central node collects information from all nodes, builds a global network view, and executes load balancing strategies when there is local congestion. When a sudden congestion is detected, the frame rate is reduced and a caching mechanism is enabled; once network conditions improve, the frame rate is gradually increased and processing performance is verified. The algorithm execution cycle includes collecting network status data every 200-500ms, making frame rate adjustment decisions every 1-2s, and performing distributed collaborative optimization every 5-10s.

[0021] In edge computing networks, the hierarchical assessment of network congestion can be achieved using the following weighted calculation model: Edge node metrics include: bandwidth utilization ( The ratio of current bandwidth usage to total bandwidth, ranging from [0,1]; packet loss rate ( The ratio of the number of lost packets to the total number of packets sent, ranging from [0,1]. Round trip delay ( ), data packet round-trip time, in milliseconds; Central node metrics include: load balancing ( The degree of load difference among edge nodes is calculated using the coefficient of variation; network topology pressure ( (a comprehensive indicator of link utilization and critical node load); The weighted calculation formula is:

[0022] in: Assign weights to each indicator, satisfying , As an indicator The normalization function maps the original values ​​to the interval [0,1].

[0023] The object detection algorithm is YOLO for real-time object detection, specifically: Load the pre-trained YOLO model and divide the entire image frame into multiple grids of equal size, with each grid responsible for predicting and identifying pedestrians; Based on each grid, multiple bounding boxes are predicted. The bounding boxes are used to enclose existing pedestrians. Each bounding box corresponds to a confidence score, which reflects both the probability of a pedestrian being inside the bounding box and the degree of fit between the bounding box and the actual location of the pedestrian. Each grid cell predicts the probability that the target within that grid cell belongs to the pedestrian category. During the prediction process, the category probability is multiplied by the confidence score of the bounding box to obtain the final score of each bounding box. The final score takes into account both the probability of the target's existence and the accuracy of the category judgment. The YOLO algorithm uses non-maximum suppression (NMS) to select high-scoring bounding boxes, suppress low-scoring bounding boxes that have a high degree of overlap with high-scoring bounding boxes, and retain the bounding boxes that accurately capture pedestrians. It identifies all pedestrians in the image frame and locates the position of each pedestrian in the image based on the finally determined bounding boxes.

[0024] The confidence level is calculated using the following formula:

[0025] in: This represents the probability that an object exists within the bounding box. To predict the intersection-union ratio (IoU) between the bounding box and the ground truth bounding box; Category conditional probability

[0026] This indicates that when a target exists within the grid, that target belongs to the category. The probability, P(⋅), is the data representation symbol for probability, expressing the probability of an event occurring. Representing category c, This indicates the existence of a target event; Final score calculation

[0027] This formula multiplies and fuses the target presence confidence of the bounding box with the category prediction accuracy, achieving a dual constraint: if... This indicates that there is a high probability that there is no target within the bounding box, regardless of... Why is it that the final scores all tend to be close to 0; if This indicates that even if a target exists within the bounding box, it belongs to the category. The probability is extremely low, and the final score is close to 0; among them : bounding box Confidence level; : bounding box Category The final score.

[0028] The detection-based tracking algorithm bytetrack performs labeling on identified pedestrians and outputs the dwell time of the target within the monitored area in the current frame. The steps are as follows: Based on previously obtained pedestrian information, bytetrack initiates the tracking process. When new image frames are received, it filters the detected pedestrian targets and distinguishes between high-confidence and low-confidence detection results. For high-confidence targets, the algorithm matches the high-confidence targets with existing tracking trajectories, calculates the similarity of appearance features and the predicted value of the motion trajectory, determines whether they are the same pedestrian, and updates the label information of the trajectory if the match is successful. For low-confidence targets, the low-confidence targets are used as candidate targets and are matched with temporarily lost trajectories. When a brand-new pedestrian target is detected and cannot be matched with any existing trajectory, a unique label is assigned to it and a new tracking trajectory is established.

[0029] This application's hierarchical matching mechanism improves tracking accuracy in complex scenes, integrates appearance and motion features, reduces ID switching rate, improves MOTA index by 8%, enhances stability through trajectory lifecycle management, reduces fragmented trajectories, increases average length by 40%, ensures accurate dwell time calculation through precise trajectory association, avoids the impact of ID jumps, optimizes efficiency through phased matching, improves calculation speed by 25%, reduces memory usage by 30%, can cope with detection fluctuations, increases recall rate in occluded scenes by 12%, adapts to multiple scenarios, and provides support for anomaly warning. Compared with similar algorithms, it is superior in ID switching rate and speed index, making it an efficient pedestrian tracking solution.

[0030] Furthermore, for each tagged pedestrian trajectory, bytetrack monitors its position in the image frame in real time. When a pedestrian first enters the monitored area, the algorithm starts timing and records the entry time. In subsequent consecutive frames, the pedestrian remains within the monitored area, and the timer continues to accumulate. If the pedestrian leaves the monitored area, the timer pauses. The accumulated time at this point is the time the pedestrian spent within the monitored area. Upon reaching the current frame, the algorithm, based on real-time location determination, outputs the corresponding pedestrian label and the cumulative dwell time within the monitored area, thus tracking and recording the pedestrian's dwelling status.

[0031] This application can accurately capture the complete dwell period of each pedestrian in the monitored area, from the initial entry to the exit, ensuring continuous and non-repeating time statistics. This provides an accurate data foundation for subsequent analysis. The real-time output of tags and cumulative time can be efficiently linked with the alarm logic module, triggering an early warning immediately upon timeout, improving the timeliness of anomaly response. Continuous tracking of dwell status can handle complex scenarios, such as when a pedestrian briefly leaves and then re-enters, the system will restart the timer to avoid time confusion. Continuous frame monitoring also reduces timing errors caused by detection fluctuations. The complete time record can be used to analyze the patterns of pedestrian flow in the area, providing data support for resource allocation and area planning, and enhancing the practicality and reliability of the system in intelligent monitoring.

[0032] The operation steps within the alarm logic module are as follows: It continuously receives real-time dwell time data of each pedestrian target in the current frame and associates it with the corresponding target label; A preset dwell time threshold is matched for each target. By comparing the cumulative dwell time of each target with the corresponding threshold frame by frame, it is determined whether the alarm condition is triggered. If the dwell time of a target exceeds the threshold for the first time, a control signal is generated to drive the alarm light to flash as a warning. For targets that have triggered alarms, continuously monitor the target status. If the dwell time continues to increase, the alarm signal remains unchanged. An alarm is triggered simultaneously when multiple targets exceed the threshold at the same time.

[0033] The further steps within the alarm logic module are as follows: When the target leaves the monitoring area, the position judgment function identifies the change in status. After confirming that the target has left, a stop signal is generated to stop the flashing of the alarm light. There are multiple alarm targets at the same time. Each target is handled independently. When a target leaves, its associated alarm signal is stopped, while other targets that are still in the area and exceed the threshold continue to keep the alarm.

[0034] The alarm logic module ignores single state transitions that cause brief misjudgments of location due to obstruction or rapid movement; if the target re-enters the area after it has completely left, the dwell time will be re-accumulated.

[0035] This application accurately associates target tags with dwell time, compares thresholds frame by frame to ensure that an alarm is triggered the first time a single target times out, and triggers multiple targets simultaneously. This ensures timely warnings and avoids omissions. Alarmed targets are continuously monitored, and the signal remains stable as long as the status remains unchanged. Multiple targets are processed independently, and the corresponding alarm stops when a single target leaves without affecting others, improving alarm accuracy and reducing invalid interference. The fault tolerance mechanism can ignore brief position misjudgments and avoid frequent alarm switching due to occlusion or rapid movement. When a target re-enters, the timer restarts to prevent time confusion and ensure timing accuracy. The entire process achieves dynamic linkage between alarms and target status, providing clear guidance for anomaly handling and retaining reliable data for subsequent analysis, enhancing the system's practicality and stability in complex monitoring scenarios.

[0036] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A real-time video stream intelligent alarm system based on platform, characterized in that, The intelligent alarm system comprises: A data acquisition module, based on a monitoring camera, reads point camera image data, and after standardization and normalization of the data, unified image data is input; A target discovery module, which inputs pre-processed image frames and monitored areas, uses a target detection-based algorithm to identify and locate pedestrians in the image frames, and uses a detection-based tracking algorithm bytrack to label the identified pedestrians, and outputs the current frame target's stay time in the monitored area; An alarm logic module, which inputs the current frame target's stay time, judges the stay time of different targets through multi-target analysis, and if it exceeds the set threshold, outputs a signal to control the alarm lamp to flash, and when the target leaves the specified area, outputs a signal to control the alarm lamp to stop working.

2. The intelligent alarm system based on real-time video stream of platform according to claim 1, characterized in that: The data standardization step converts the color image into a grayscale image, reduces the data dimension and unifies the color space, calculates the pixel mean of the image data set, subtracts the mean from each pixel value to move the data distribution center to 0 point, divides the centralized data by the standard deviation to make the standard deviation of the data 1, and completes the standardization processing; The normalization processing adjusts the size, converts the color format, and converts it to a floating-point type; the pixel value is scaled from 0-255 to the [0, 1] interval.

3. The intelligent real-time video stream alarm system based on platform according to claim 1, characterized in that: The pre-processing uses a distributed multi-thread processing method for video streams, dynamically adjusts the processing frame rate according to the camera network congestion through a distributed ABR algorithm, and the specific steps of the ABR algorithm are: Network state monitoring, edge nodes collect local bandwidth utilization, packet loss rate and round-trip time data in real time, center nodes statistically analyze global load balancing state and network topology, based on the monitoring data, the network congestion level is evaluated, and the congestion score is calculated by weighting; Based on the congestion score, dynamically adjust the video processing frame rate, define three congestion levels of low, medium and high, corresponding to different recommended frame rate ranges and processing strategies; The center node and the edge node cooperatively execute global optimization, the edge node adjusts the frame rate based on the local network state and reports the state, the center node collects all node information, constructs a global network view, and executes a load balancing strategy when there is local congestion; When a sudden congestion is detected, the frame rate is reduced and the cache mechanism is enabled; after the network condition improves, the frame rate is gradually increased and the processing performance is verified; The algorithm execution cycle includes collecting network state data every 200-500ms, executing frame rate adjustment decision every 1-2s, and performing distributed collaborative optimization every 5-10s.

4. The intelligent real-time video stream alarm system based on platform according to claim 1, characterized in that, The target detection algorithm is yolo real-time target detection, specifically: Load the pre-trained yolo model, divide the entire image frame into multiple equal-sized grids, and each grid is responsible for predicting and identifying pedestrians; Based on each grid, multiple bounding boxes are predicted, the bounding box is used to frame the existing pedestrians, and each bounding box corresponds to a confidence score, which reflects the possibility of the existence of pedestrians in the bounding box and the degree of fit between the bounding box and the actual position of the pedestrian; Each grid predicts the probability that the target in the grid belongs to the pedestrian category, and the category probability is multiplied by the confidence of the bounding box during the prediction process to obtain the final score of each bounding box, which considers the possibility of the existence of the target and the accuracy of the category judgment; The YOLO algorithm uses the non-maximum suppression (NMS) technique to filter out high-score bounding boxes, suppress low-score bounding boxes with high overlap with high-score bounding boxes, and retain accurate bounding boxes that enclose pedestrians. All pedestrians in the image frame are identified based on the final determined bounding boxes to locate the position of each pedestrian in the image.

5. The intelligent real-time video streaming based platform alarm system according to claim 1, wherein, The detection-based tracking algorithm bytetrack processes the identified pedestrians with labels and outputs the staying time of the target in the monitoring area in the current frame as follows: The bytetrack relies on the pedestrian information obtained in the early stage to start the tracking process. When a new image frame is input, the detected pedestrian targets are filtered to distinguish high-confidence and low-confidence detection results. For high-confidence targets, the algorithm matches the high-confidence target with the existing tracking trajectory, calculates the appearance feature similarity and motion trajectory prediction value, determines whether it is the same pedestrian, and updates the label information of the trajectory if the matching is successful. For low-confidence targets, the low-confidence target is taken as a candidate target and is matched with a temporarily lost trajectory. When a completely new pedestrian target is detected and cannot be matched with any existing trajectory, a unique label is assigned to it to establish a new tracking trajectory.

6. The intelligent real-time video streaming based platform alarm system according to claim 5, characterized in that: Further, for each labeled pedestrian trajectory, the bytetrack monitors its position in the image frame in real time. When the pedestrian first enters the monitoring area, the algorithm starts timing and records the entry time. In the subsequent continuous frames, if the pedestrian is still in the monitoring area, the timing continues to accumulate. If the pedestrian leaves the monitoring area, the timing is paused, and the accumulated time is the staying time of the pedestrian in the monitoring area. In the current frame, the algorithm outputs the corresponding pedestrian label and the accumulated staying time in the monitoring area based on the real-time position judgment, tracks and records the staying state of the pedestrian.

7. The intelligent real-time video streaming based platform alarm system according to claim 1, wherein, The operation steps in the alarm logic module are as follows: Real-time staying time data of each pedestrian target in the current frame is continuously received and associated with the corresponding target label one by one. A preset staying time threshold is matched for each target. By comparing the accumulated staying time of each target with the corresponding threshold frame by frame, it is determined whether the alarm condition is triggered. If the staying time of a target exceeds the threshold for the first time, a control signal is generated to drive the alarm light into a flashing state for prompting. For the target that has triggered the alarm, the target state is continuously monitored. If the staying time continues to increase, the alarm signal remains unchanged. If multiple targets exceed the threshold at the same time and trigger the alarm simultaneously.

8. The intelligent real-time video stream alarm system based on platform according to claim 1, characterized in that: The further operation steps in the alarm logic module are as follows: When the target leaves the monitoring area, the state change is identified by a position judgment function, and a stop signal is generated to terminate the flashing state of the alarm light after confirming that the target has left. If there are multiple alarm targets at the same time, each target is processed independently. When one target leaves, the corresponding alarm signal is stopped, and the other targets that are still in the area and exceed the threshold continue to maintain the alarm.

9. The intelligent real-time video streaming based platform alarm system according to claim 1, wherein: The alarm logic module ignores single state jump for short-term position misjudgment caused by shielding and rapid movement, and accumulates the stay time again when the target enters the area again after leaving the area completely.

Images