Methods, systems, devices and storage media for passenger fall warning in rail transit stations
By acquiring key points of passenger skeletons in rail transit stations and combining them with a regional semantic parameter library for differentiated weighting and two-level verification, the problems of insufficient regional adaptability and high false alarm rate in passenger fall detection are solved, achieving accurate perception and intelligent early warning, reducing the false alarm rate and simplifying engineering implementation.
Patent Information
- Application Number
- CN202611241588.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-10
AI Technical Summary
Existing passenger fall detection technologies in rail transit stations suffer from insufficient regional adaptability, high false alarm rates, delayed warnings, and difficulties in implementation. In particular, the differences in fall characteristics across different functional areas are not fully utilized, resulting in low recall rates in high-risk areas and high false alarm rates in low-risk areas.
By acquiring station surveillance video streams, extracting key points of passengers' human skeletons, combining them with a regional semantic parameter library for differentiated weighted calculation and two-level false alarm verification, and using scene object semantic reasoning to accurately filter normal behavior and output graded alarm information.
It achieves accurate perception in different functional areas, reduces the false alarm rate to less than 7%, improves the detection recall rate, simplifies the engineering implementation cost, and realizes intelligent early warning and rapid response.
Smart Images

Figure CN123072170A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of safety monitoring technology for rail transit stations, and in particular to a method, system, device, and storage medium for early warning of passenger falls in rail transit stations. Background Technology
[0002] With the continuous increase in passenger flow density in urban rail transit, passenger falls have become a frequent safety incident at stations, which can easily lead to secondary accidents such as stampedes and congestion, affecting operational safety and service quality.
[0003] Currently, fall safety monitoring at train stations mainly falls into two categories of technical solutions. The first category is the manual monitoring solution, which relies on on-site inspections by station staff and manual monitoring at the control center. This solution has limitations such as limited monitoring coverage, easy omissions during peak hours, and inability to provide advance warnings.
[0004] The second category consists of automatic detection schemes based on general human posture recognition. These schemes acquire video streams from station surveillance cameras, use a general posture estimation model to extract key points of the human skeleton, and construct judgment rules based on geometric features such as torso tilt angle and changes in center of gravity height. The schemes then output fall detection results and trigger alarms. Some schemes improve detection performance by optimizing the network structure or adding multimodal hardware such as millimeter-wave radar and pressure sensors.
[0005] However, existing automatic detection solutions generally use a uniform feature weight and judgment threshold across the entire domain, without making differentiated designs for the different fall characteristics in different functional areas such as escalators, stairs, and rest areas. This results in an inability to balance the detection recall rate in high-risk areas with the false alarm rate in low-risk areas. In addition, existing pure vision solutions rely solely on the geometric features of the human skeleton for judgment, without combining them with fixed objects in the scene for logical verification. The skeletal features of normal behaviors such as bending over to pick up items, sitting down to rest, and organizing luggage in the station are highly similar to the fall posture, which can easily lead to false alarms. Summary of the Invention
[0006] In view of this, the purpose of this invention is to overcome the shortcomings of the prior art and provide a method, system, device and storage medium for passenger fall early warning in rail transit stations, aiming to solve four major technical problems in fall detection in rail transit stations: insufficient regional adaptability, high false alarm rate, delayed early warning and difficulty in implementation.
[0007] To achieve the above objectives, the technical solution of the present invention is as follows: The first objective of this invention is to provide a passenger fall warning method for rail transit stations, comprising the following steps: Acquire station surveillance video stream, extract the coordinates of key points of the human skeleton of passengers in the video frames, and generate skeleton data; Obtain a semantic parameter library for the station area, which includes the coordinate mapping relationship of each functional area, as well as the set of attitude feature weights and risk judgment thresholds corresponding to each functional area. Based on the passenger's location, a corresponding functional area is matched, and the posture feature weight set corresponding to that functional area is called to perform weighted calculation on the skeletal data to obtain an initial risk value. For passengers whose initial risk value exceeds the risk judgment threshold of the corresponding functional area, a false alarm check is performed. The false alarm check includes correcting the initial risk value based on the regional attributes of the functional area where the passenger is located, and / or performing behavioral reasoning based on the spatial positional relationship between objects in the passenger's surrounding scene and key points of the human skeleton to correct the risk value and output a valid risk value. Alarm information is generated and output based on the effective risk value.
[0008] In some implementations, the false alarm verification further includes: Level 1 verification: The initial risk value is attenuated and corrected based on the regional attributes of the functional area where the passenger is located; Level 2 verification: Identify scene objects within a preset range around the passenger, determine the passenger's behavior type based on the spatial relationship between key points of the human skeleton and the scene objects, and correct the risk value based on the behavior type. In the second level of verification: When the spatial relationship between the key points of the human buttocks and the seat meets the preset sitting judgment conditions, it is judged as a normal sitting behavior and the risk value is cleared to zero. When the spatial relationship between the key points of a person's hand and the object dropped on the ground meets the preset bending judgment conditions, it is judged as bending down to pick up the object, and the risk value is reduced by the first proportion. When the spatial relationship between the human body and the suitcase meets the preset sorting judgment conditions, it is judged as bending over to sort luggage behavior, and the risk value is reduced by the second proportion.
[0009] In some implementations, it also includes: Obtain the skeletal temporal sequence of multiple consecutive frames; Extract gait abnormality precursor features from the skeletal time sequence; The gait abnormality precursor features are input into a temporal neural network for encoding, and a preliminary risk score is output. When the pre-risk score exceeds the warning threshold, a pre-warning signal is generated.
[0010] In some implementations, the gait abnormality precursor features include at least one of the following: stride abrupt change rate, lateral sway amplitude of the center of gravity, duration of trunk tilt, and degree of gait continuity interruption. Different functional areas are configured with different warning thresholds, and the warning threshold for high-risk functional areas is lower than that for low-risk functional areas.
[0011] In some implementations, the functional areas are divided into high-risk areas, medium-risk areas, and low-risk areas according to their security risk levels; The high-risk area is configured with a first weight set and a first judgment threshold. In the first weight set, the weight of the trunk tilt feature is greater than the weight of the limb extension feature, and the first judgment threshold is lower than the benchmark threshold, so that the high-risk area is more sensitive to abnormal trunk tilt. The medium-risk area is configured with a second weight set and a second judgment threshold, and the weights of each posture feature in the second weight set are evenly distributed. The low-risk area is configured with a third weight set and a third judgment threshold. In the third weight set, the weight of the limb extension feature is greater than the weight of the trunk tilt feature, and the third judgment threshold is higher than the benchmark threshold, so that the low-risk area has a lower false alarm rate.
[0012] In some implementations, generating and outputting alarm information includes: Risk levels are determined based on the magnitude of the effective risk value and / or the importance of the functional area where the passenger is located; Match the corresponding alarm output method according to the risk level; The system outputs tiered alarm information to the station operation system through a standardized interface.
[0013] In some embodiments, the linkage monitoring and early warning system further includes a drone inspection module and an image recognition unit; the drone inspection module is used to collect slope image data and transmit it to the image recognition unit; the image recognition unit is used to identify slope cracks, gullies and water accumulation areas from the images, and the identification results are transmitted to the data processing unit to assist in the judgment of slope stability and early warning decision-making.
[0014] In some embodiments, a typhoon protection component is also included, which includes a flexible windbreak located on the upper part of the slope and stepped guide channels spaced apart along the slope.
[0015] A second objective of this invention is to provide a passenger fall warning system for rail transit stations, comprising: The video acquisition module is used to acquire the station's surveillance video stream; The human pose extraction module is used to extract the coordinates of key points of the passenger's human skeleton from the video stream and generate skeletal data; The region semantic labeling module is used to store the coordinate mapping relationship of each functional region, as well as the set of pose feature weights and risk judgment thresholds corresponding to each functional region; The risk calculation module is used to match the corresponding functional area based on the passenger's location, call the corresponding posture feature weight set to perform weighted calculation on the skeletal data, and obtain an initial risk value. The false alarm verification module is used to correct the risk value of passengers whose initial risk value exceeds the corresponding risk judgment threshold based on the regional attributes and / or the spatial positional relationship between scene objects and key points of the human skeleton, and output a valid risk value. The alarm output module is used to generate and output alarm information based on the effective risk value; The temporal risk warning module is used to acquire a continuous multi-frame skeletal temporal sequence, extract gait abnormality precursor features from the skeletal temporal sequence, input the gait abnormality precursor features into a temporal neural network for encoding, output a preliminary risk score, and generate a preliminary warning signal when the preliminary risk score exceeds a warning threshold.
[0016] In some implementations, the alarm output module classifies risk levels based on the magnitude of the effective risk value and / or the importance of the functional area where the passenger is located, matches the corresponding alarm output method according to the risk level, and outputs graded alarm information to the station operation system through a standardized interface.
[0017] A third objective of this invention is to provide a passenger fall warning device for rail transit stations, comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method described above.
[0018] A fourth objective of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.
[0019] Compared with the prior art, the present invention has at least the following beneficial effects: The technical solution of this invention, through regional semantic labeling, differentiated weighted calculation, two-level false alarm verification, and hierarchical alarm output, addresses the issue of differentiated risk perception in different regions. The regional semantic parameter library and differentiated weighted calculation work synergistically, enabling the same posture to generate different risk quantification results in different functional areas. High-risk areas (escalators / stairs) are more sensitive to abnormal postures due to high weighting for torso tilt angle and low judgment thresholds, while low-risk areas (rest areas / corridors) effectively filter out normal sitting and lying postures due to high weighting for limb extension and high judgment thresholds. This fundamentally solves the problem of the existing technology's overall risk assessment across all regions. The model addresses the dilemma of strict threshold for missed detections and lenient threshold for false alarms. In terms of pure visual semantic verification, a two-level verification mechanism, using pre-attenuation of regional prior knowledge and semantic reasoning of scene objects, accurately filters normal behaviors such as "sitting down, bending over to pick up items, and organizing luggage," reducing the overall false alarm rate from 18% to below 7%, all without requiring additional hardware sensors. Regarding tiered operation, effective risk values are converted into three-level alarms matching station operation procedures, directly connecting to broadcasting, video surveillance, and work order systems through standardized interfaces, reducing secondary development workload by more than 60%. These five steps work together to form a complete technical chain, achieving the overall technical effect of "accurate perception, intelligent verification, proactive early warning, and rapid implementation," effectively solving all the core technical problems pointed out in the background: insufficient regional adaptability, high false alarm rate, and disconnect between alarms and operations. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating the passenger fall warning method for rail transit stations in an embodiment of the present invention. Figure 2 This is a block diagram of the overall architecture of the passenger fall warning system in a rail transit station according to an embodiment of the present invention; Figure 3 This is a block diagram of the overall architecture of the passenger fall warning system in a rail transit station according to an embodiment of the present invention; Figure 4 This is a structural block diagram of a passenger fall warning system for rail transit stations in an embodiment of the present invention; Figure 5 This is a structural block diagram of a passenger fall warning device for rail transit stations in an embodiment of the present invention.
[0021] Explanation of reference numerals in the attached figures: 10-Video capture module; 20 - Human posture extraction module; 30 - Region semantic labeling module; 40 - Risk Calculation Module; 50 - False Alarm Detection Module; 51 - Region Pre-verification Unit; 52 - Object Context Verification Unit; 60 - Alarm Output Module; 100 - Processor; 200 - Storage device. Detailed Implementation
[0022] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] In the description of this invention, it should be noted that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0024] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can also refer to the internal connection of two components; and they can refer to a wireless connection or a wired connection. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0025] refer to Figures 1-3 As shown, one embodiment of the present invention provides a passenger fall warning method for rail transit stations, the fall warning method comprising the following steps: S1: Obtain the station monitoring video stream, extract the coordinates of key points of the human skeleton of passengers in the video frames, and generate skeleton data; S2: Obtain the semantic parameter library of the station area, which includes the coordinate mapping relationship of each functional area, as well as the set of attitude feature weights and risk judgment thresholds corresponding to each functional area; S3: Match the corresponding functional area based on the passenger's location, call the posture feature weight set corresponding to the functional area to perform weighted calculation on the skeleton data, and obtain the initial risk value; S4: For passengers whose initial risk value exceeds the risk judgment threshold of the corresponding functional area, perform false alarm verification. The false alarm verification includes correcting the initial risk value according to the regional attributes of the functional area where the passenger is located, and / or performing behavioral reasoning based on the spatial positional relationship between objects in the passenger's surrounding scene and key points of the human skeleton to correct the risk value and output a valid risk value. S5: Generate and output alarm information based on the effective risk value.
[0026] Thus, by using a regional semantic parameter library and differentiated weighted calculation, adaptive risk perception of "same posture, different regions, different judgments" is achieved. High torso tilt weight and low judgment threshold are configured for high-risk areas such as escalators / stairs, improving detection sensitivity; high limb extension weight and high judgment threshold are configured for low-risk areas such as rest areas, reducing false alarms of normal sitting and lying down, and solving the dilemma of not being able to balance recall and false alarm rate in a unified model across the entire domain; through a two-level false alarm verification mechanism, combining regional attribute pre-attenuation (low-cost coarse filtering) and scene object semantic reasoning (high-precision fine verification), the system uses prior knowledge of scenes such as seats, fallen objects, and suitcases to accurately distinguish between normal behaviors such as sitting down, bending over to pick up objects, and organizing luggage and falling behaviors, effectively reducing false alarms without the need for additional hardware sensors; through hierarchical alarm output, effective risk values are converted into three-level alarms that match the station's operation process, and the system directly connects to the station's existing system through standardized interfaces, reducing the cost of engineering implementation.
[0027] refer to Figure 3 As shown, in one embodiment of the present invention, the false alarm verification further includes: Level 1 verification: Obtain the area type label of the functional area where the passenger is located. When the area type label belongs to the preset low-risk area type, the initial risk value is attenuated according to the preset first attenuation ratio to obtain the first corrected risk value.
[0028] Therefore, the first-level verification utilizes prior knowledge that passengers in low-risk areas (such as rest areas) have a high probability of sitting or lying down normally. With extremely low computational cost (table lookup combined with multiplication operations), it pre-attenuates the risk value, quickly eliminating a large number of primary suspected targets without running any detection model. This significantly reduces the amount of data entering the second-level verification and saves computing resources at the edge.
[0029] Second-level verification: When the first corrected risk value or the initial risk value exceeds the risk judgment threshold of the corresponding functional area, identify scene objects within a preset range around the passenger. The scene objects include at least seats and / or ground objects. Calculate the spatial positional relationship between the key points of the human skeleton and the scene objects. Determine the passenger behavior type based on the spatial positional relationship. The behavior type includes normal behavior type and suspected fall type. When determined to be a normal behavior type, attenuate or clear the risk value according to a preset second attenuation ratio and output a valid risk value. When determined to be a suspected fall type, maintain the current risk value as a valid risk value.
[0030] The second-level verification incorporates semantic information from scene objects (seats, fallen objects, etc.) to logically fuse pure visual posture features ("human body in a low posture") with scene context ("seat under buttocks" / "fallen object nearby"). When the human body is in a low posture and supported by a seat, it is judged as normal sitting and the risk value is cleared to zero. When the hand is close to an object on the ground, it is judged as bending down to pick up the object and the risk value is significantly reduced. This fundamentally achieves the semantic distinction between "actively squatting / sitting" and "passively falling", rather than relying solely on the similarity comparison of skeletal features.
[0031] refer to Figure 3 As shown, in one embodiment of the present invention, the second-level verification process, which determines the behavior as normal and corrects the risk value, includes: When the vertical projection of the center point of the human hip is within the horizontal range of the seat, and the vertical distance between the center point of the human hip and the upper surface of the seat is less than the first preset distance threshold, it is determined to be a normal sitting behavior, and the current risk value is cleared to zero. When the horizontal distance between the key point of the human wrist and the object falling on the ground is less than the second preset distance threshold, and the vertical height of the key point of the wrist is lower than the vertical height of the center point of the human hip, it is determined to be a bending-over picking up behavior, and the current risk value is attenuated according to the preset second attenuation ratio. When the horizontal distance between the center point of the human hip and the suitcase is less than the third preset distance threshold, and the tilt angle of the human torso is greater than the preset angle threshold, it is determined to be a bending-over-to-organize-luggage behavior, and the current risk value is attenuated according to the preset third attenuation ratio.
[0032] In this embodiment, the first preset distance threshold is 0.2 meters, the second preset distance threshold is 0.3 meters, the third preset distance threshold is 0.5 meters, and the preset angle threshold is 30 degrees; the second attenuation ratio is 70%, and the third attenuation ratio is 60%.
[0033] In one embodiment of the present invention, a skeletal temporal sequence of N consecutive frames is obtained, where N is an integer greater than 1; Gait abnormality precursor features are extracted from the skeletal time sequence. The gait abnormality precursor features include at least one of the following: stride abrupt change rate, lateral sway amplitude of center of gravity, duration of trunk tilt, and degree of gait continuity interruption. The gait abnormality precursor features are input into a temporal neural network for encoding, and the temporal neural network outputs a preliminary risk score, which is used to characterize the degree of risk of a passenger falling in the future time period. When the pre-risk score exceeds a preset warning threshold, a pre-warning signal is generated.
[0034] It should be further explained that the preferred implementation of the temporal neural network is TCN (Temporal Convolutional Network). TCN employs a dilated convolution architecture, which can achieve a large temporal receptive field with a small number of parameters—capable of covering a continuous skeletal temporal sequence of 15 frames (approximately 0.6 seconds). By learning the evolution trend of gait features over the past N frames, TCN establishes a mapping relationship between "gait deterioration pattern → future fall risk," thereby enabling it to output a preliminary risk score higher than the warning threshold 1 to 3 seconds before the actual fall occurs.
[0035] Specifically, during the training phase, the TCN model uses sample data labeled with the "time of the fall" to learn and identify abnormal gait evolution patterns within a time window preceding the fall. During the inference phase, the model receives continuous skeletal sequences prior to the current moment and outputs a prediction score for future fall risk based on the learned temporal patterns. This "predicting future states based on historical sequences" approach is a key difference from existing post-event detection schemes that "determine the current state based on the current frame."
[0036] In one embodiment of the present invention, the temporal neural network is a temporal convolutional network; Temporal convolutional networks consist of multiple cascaded temporal convolutional blocks, each of which includes dilated convolutional layers, batch normalization layers, and activation function layers. Each dilated convolutional layer in a temporal convolutional block has a different dilation rate. Along the direction from the input layer to the output layer of the temporal convolutional network, the dilation rate of each dilated convolutional layer increases layer by layer. The temporal receptive field of the temporal convolutional network covers the skeletal temporal sequence of N consecutive frames. The temporal receptive field is the range of historical frames used by the temporal convolutional network when performing one forward inference. The temporal convolutional network outputs a preliminary risk score based on the skeletal temporal sequence within the temporal receptive field. The output of the temporal convolutional network is connected to a fully connected layer. The fully connected layer maps the feature vector output by the temporal convolutional network to a single numerical value. After being mapped to the 0 to 1 range by an activation function, the pre-risk score is output.
[0037] In one embodiment of the present invention, the preset warning threshold is configured according to the regional risk level of the functional area where the passenger is located, including: Obtain the regional risk level of the functional area where the passenger is currently located. The regional risk level is predetermined based on the type of functional area and / or historical fall accident statistics. The corresponding early warning threshold is matched according to the regional risk level. The functional area with the first regional risk level is configured with the first early warning threshold, and the functional area with the second regional risk level is configured with the second early warning threshold. The first regional risk level is higher than the second regional risk level, and the first early warning threshold is lower than the second early warning threshold. A pre-warning signal is generated when the pre-risk score exceeds the warning threshold. The value of the warning threshold determines the sensitivity of triggering the pre-warning. The lower the warning threshold, the more likely passengers are to trigger the pre-warning signal with a lower pre-risk score, thus enabling functional areas with higher risk levels to have higher warning sensitivity.
[0038] In one embodiment of the present invention, the functional area is divided into multiple area types according to its security risk level, and the multiple area types include high-risk area type, medium-risk area type and low-risk area type; The set of posture feature weights includes trunk tilt angle weight, center of gravity height change rate weight, and limb extension weight. The high-risk area type is configured with a first weight set and a first judgment threshold. In the first weight set, the weight of the trunk tilt angle is greater than at least one of the weight of the rate of change of the center of gravity height and the weight of the limb extension. The value of the first judgment threshold is the first value. The medium-risk area type is configured with a second weight set and a second judgment threshold. In the second weight set, the absolute value of the difference between the trunk tilt angle weight, the center of gravity height change rate weight and the limb extension weight is less than the first preset difference. The value of the second judgment threshold is the second value. The low-risk area type is configured with a third weight set and a third judgment threshold. In the third weight set, the weight of limb extension is greater than at least one of the weight of trunk tilt angle and the weight of the rate of change of center of gravity height. The value of the third judgment threshold is the third value. The first determination threshold is less than the second determination threshold, and the second determination threshold is less than the third determination threshold.
[0039] See Figure 4 As shown, another embodiment of the present invention also provides a passenger fall warning system for rail transit stations. The system includes a video acquisition module 10, a human posture extraction module 20, a region semantic labeling module 30, a risk calculation module 40, a false alarm verification module 50, and an alarm output module 60, wherein: Video acquisition module 10 is used to acquire station monitoring video streams; The human pose extraction module 20 is connected to the video acquisition module 10 and is used to extract the coordinates of key points of the human skeleton of the passenger in the video frame from the video stream and generate skeleton data. The region semantic calibration module 30 is used to store the region semantic parameter library, which contains the coordinate mapping relationship of each functional region, as well as the set of pose feature weights and risk judgment thresholds corresponding to each functional region. The risk calculation module 40 is connected to the human posture extraction module 20 and the region semantic labeling module 30 respectively. It is used to match the corresponding functional area according to the passenger's location, call the posture feature weight set corresponding to the functional area to perform weighted calculation on the skeletal data, and obtain the initial risk value. The false alarm verification module 50 is connected to the risk calculation module 40 and is used to perform false alarm verification for passengers whose initial risk value exceeds the risk judgment threshold of the corresponding functional area. The false alarm verification includes correcting the initial risk value according to the regional attributes of the functional area where the passenger is located, and / or performing behavioral reasoning based on the spatial positional relationship between objects in the passenger's surrounding scene and key points of the human skeleton to correct the risk value and output a valid risk value. The alarm output module 60 is connected to the false alarm verification module 50 and is used to generate and output alarm information based on the effective risk value.
[0040] More specifically, the false alarm verification module 50 includes a region pre-verification unit 51 and an object context verification unit 52, wherein: The area pre-verification unit 51 is used to obtain the area type label of the functional area where the passenger is located. When the area type label belongs to the preset low-risk area type, the initial risk value is attenuated according to the preset first attenuation ratio, and the first corrected risk value is output. The object context verification unit 52 is connected to the area pre-verification unit 51. When the first corrected risk value or the initial risk value exceeds the risk judgment threshold of the corresponding functional area, it identifies scene objects within a preset range around the passenger, calculates the spatial positional relationship between the key points of the human skeleton and the scene objects, and determines the passenger behavior type based on the spatial positional relationship. The behavior type includes normal behavior type and suspected fall type. When it is determined to be a normal behavior type, the risk value is attenuated or cleared according to the preset second attenuation ratio, and a valid risk value is output. When it is determined to be a suspected fall type, the current risk value is maintained as a valid risk value.
[0041] like Figure 5 As shown, one embodiment of the present invention provides a passenger fall warning device for rail transit stations. The device includes one or more processors 100 and a storage device 200 for storing one or more programs.
[0042] The processor 100 can be a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), a Neural Processing Unit (NPU), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices. In this embodiment, the processor 100 preferably adopts a heterogeneous computing architecture, including a CPU and an NPU, where the CPU is responsible for logic control and data scheduling, and the NPU is responsible for accelerating inference of the human pose estimation model and the temporal neural network. Due to the limited power consumption and heat dissipation conditions of the edge computing nodes in rail transit stations, the selection of the processor 100 needs to balance computing power and power consumption, and a low-power embedded AI processor with a computing power of 4 TOPS or higher can be used.
[0043] Storage device 200 may be read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), solid state drive (SSD), or other forms of non-transitory computer-readable storage media. Storage device 200 is used to store one or more programs, as well as intermediate data (including skeletal data, region semantic parameter libraries, initial risk values, effective risk values, pre-risk scores, etc.) and runtime logs generated during program execution.
[0044] When one or more programs are executed by the one or more processors 100, the one or more processors perform the following steps: Step 1: Acquire the station's surveillance video stream, extract the coordinates of key points of the passenger's human skeleton in the video frames, and generate skeletal data. Specifically, the processor 100 receives the real-time video stream collected by the station's existing surveillance cameras through a data interface (such as a gigabit Ethernet interface), decodes the video stream and extracts frames, calls a lightweight pose estimation model (such as YOLOv8n-pose) stored in the storage device 200 to perform inference on each frame of the video image, and outputs the two-dimensional coordinates and confidence scores of 17 key skeletal points of the human body, forming structured skeletal data.
[0045] In a preferred implementation of this embodiment, the processor 100 also executes a target tracking algorithm (such as Kalman filtering combined with Hungarian matching) to perform cross-frame identity association for the same passenger between adjacent frames, ensuring that the skeletal data of each passenger is continuous in the time dimension, thereby generating a skeletal temporal sequence of the same passenger in multiple consecutive frames.
[0046] Step 2: Obtain the station area semantic parameter library. The storage device 200 pre-stores the station area semantic parameter library, which includes the coordinate mapping relationships of each functional area, as well as the corresponding posture feature weight sets and risk assessment thresholds for each functional area. Specifically, the functional areas are divided into high-risk areas (including escalator areas and stair areas), medium-risk areas (including platform areas and turnstile areas), and low-risk areas (including pedestrian walkways and seating areas) based on the station's spatial layout and safety risk level. The weight sets and risk assessment thresholds for each area are pre-calibrated and stored in the storage device based on historical fall accident statistics and expert experience.
[0047] Step 3: Match the corresponding functional area based on the passenger's location, and use the posture feature weight set corresponding to that functional area to perform weighted calculation on the skeletal data to obtain an initial risk value. Specifically, the processor 100 queries the semantic parameter library of the region based on the coordinates of the passenger's hip center point to determine the functional area where the passenger is currently located, and reads the posture feature weight set and risk judgment threshold corresponding to that region; calculates the quantified values of three types of core posture features from the skeletal data: trunk vertical tilt angle, rate of change of center of gravity height, and limb extension; and performs a weighted summation of the three types of posture feature values according to the weight set to obtain an initial risk value in the range of 0 to 1.
[0048] Step 4: For passengers whose initial risk value exceeds the risk assessment threshold of the corresponding functional area, perform false alarm verification. False alarm verification includes two levels of verification: The first level of verification involves correcting the initial risk value based on the regional attributes of the functional area where the passenger is located. Specifically, when the passenger's area is a preset low-risk area type (such as a rest seating area), the processor attenuates the initial risk value according to a preset first attenuation ratio (such as 40%) to obtain a first corrected risk value.
[0049] The second level of verification involves inferring behavior and correcting the risk value based on the spatial relationship between objects in the passenger's surrounding environment and key points on the human skeleton. Specifically, when the first corrected risk value or the initial risk value still exceeds the risk judgment threshold of the corresponding functional area, the processor 100 calls a lightweight object detection model (such as YOLOv8n) to identify objects in the passenger's surrounding environment within a preset range (e.g., 1.5 meters). These objects include at least one of seats, objects dropped on the ground, and suitcases. The processor calculates the spatial relationship (including horizontal and vertical distances) between the key points on the human skeleton and the objects in the environment, and determines the passenger's behavior type based on this relationship: when it is determined to be a normal sitting behavior, the risk value is reset to zero; when it is determined to be a bending-over-to-pick-up behavior, the risk value is reduced by 70%; when it is determined to be a bending-over-to-organize-luggage behavior, the risk value is reduced by 60%; when it is determined to be a suspected fall type, the current risk value is maintained; and finally, a valid risk value is output.
[0050] Step 5: Generate and output alarm information based on the effective risk value. Specifically, the processor classifies the risk level according to the magnitude of the effective risk value and the regional importance of the functional area where the passenger is located—low risk (e.g., 0.6 to 0.7), medium risk (e.g., 0.7 to 0.85), and high risk (e.g., 0.85 and above); matches the corresponding alarm output method according to the risk level—low risk only records logs in the background, medium risk pushes to the mobile terminal of station staff, and high risk triggers video pop-ups and voice broadcasts; and outputs the graded alarm information to the station's existing operation system through a standardized interface.
[0051] In another preferred implementation of this embodiment, the processor performs the following steps simultaneously or after executing step one (i.e., implementing the advance warning method as described in any one of claims 4 to 6): Obtain the skeletal timing sequence of N consecutive frames, where N is an integer greater than 1 (preferably, N=15, corresponding to approximately 0.6 seconds). Gait abnormality precursor features are extracted from the skeletal time sequence. These features include at least one of stride abrupt change rate, lateral sway amplitude of the center of gravity, duration of trunk tilt, and degree of gait continuity interruption, preferably including all four types of features. Specifically: the stride abrupt change rate is calculated based on the ratio of the standard deviation to the mean of multiple consecutive stride lengths and is used to characterize the stability of gait rhythm; the lateral sway amplitude of the center of gravity is calculated based on the ratio of the maximum horizontal offset range of the human body's center of gravity within multiple consecutive frames to the width of the human body bounding box and is used to characterize the human body's balance control ability; the duration of trunk tilt is calculated based on the cumulative time when the trunk tilt angle continuously exceeds a preset angle threshold and is used to characterize the persistence of posture loss of control; the degree of gait continuity interruption is calculated based on the edit distance between the actual gait phase transfer sequence and the standard gait phase transfer sequence and is used to characterize the degree of abnormal deviation of the gait pattern. The gait abnormality precursor features are combined into a multi-dimensional feature vector, which is then input into a temporal neural network for encoding. The temporal neural network outputs a preliminary risk score, which is used to characterize the degree of risk of a passenger falling within a preset future time period. When the pre-risk score exceeds a preset warning threshold, the processor generates a pre-warning signal.
[0052] Furthermore, different functional areas are configured with different warning thresholds. The warning threshold for high-risk areas is lower than that for low-risk areas, making high-risk areas more sensitive to warnings. Specifically, the processor matches the corresponding warning threshold according to the regional risk level of the functional area where the passenger is located. When the pre-risk score exceeds the warning threshold, a pre-warning signal is generated.
[0053] In this embodiment, when the processor performs step five, it also generates and outputs alarm information based on the effective risk value and / or the advance warning signal. That is, it can trigger an alarm based solely on the effective risk value, or it can trigger an alarm based solely on the advance warning signal, or it can trigger an alarm by combining both (such as triggering a higher-level alarm when both exceed their respective thresholds, or triggering a corresponding-level alarm when either exceeds its threshold).
[0054] The device can be deployed on station edge computing nodes and connected to existing station surveillance cameras and existing station operation systems (including broadcasting systems, video surveillance platforms, and station work order systems) via data interfaces. Based on the limited computing resources of edge computing nodes, the processor employs optimization strategies such as lightweight model selection, on-demand triggered inference (the scene object detection model is triggered only after a primary suspected target is generated, rather than executing every frame), and model quantization (FP16 half-precision quantization) to ensure that the total end-to-end processing time for a single video stream is ≤30ms, meeting the requirements for real-time safety monitoring of rail transit stations.
[0055] In a specific application scenario of this embodiment: the device is deployed at an edge computing node in the concourse of a subway station, connecting to three surveillance cameras. When a passenger in the escalator area exhibits an abnormal posture of misstepping and continuously leaning backward, the processor executes step one to extract the passenger's skeletal data, and step two to obtain the weight set of the escalator area in the regional semantic parameter library (torso tilt angle weight 0.4, gait continuity weight 0.3, center of gravity height change rate weight 0.3) and the judgment threshold (0.6), and step three to calculate the initial risk value of 0.74, which exceeds the escalator area judgment threshold and is marked as a primary suspected target; step four is executed for the first level of verification—the passenger is located in the escalator area (high-risk area), so no attenuation is performed, and the risk value is maintained at 0.74; step four is executed for the second level of verification—calling the scene object detection model to identify the absence of surrounding seats, fallen objects, suitcases, and other scene objects. The body exhibits no matching normal behavior, classifying it as a suspected fall and maintaining a risk value of 0.74 as the effective risk value. Simultaneously, the processor extracts 15 consecutive frames of the passenger's skeletal temporal sequence, extracting gait abnormality precursor features—stride abrupt change rate 0.28, center of gravity lateral sway amplitude 0.18, trunk tilt duration 0.6 seconds, and gait continuity interruption degree 0.65. These features are input into a temporal convolutional network (TCN), outputting a pre-risk score of 0.78, exceeding the escalator area warning threshold (0.6), generating a pre-warning signal. Step five is executed: based on the effective risk value of 0.74 and the pre-warning signal of 0.78, it is classified as a medium risk, and an alarm message is pushed to the station staff's mobile terminal, achieving an advance warning approximately 2 seconds before a fall.
[0056] The early warning device in this embodiment executes a program stored in a storage device through a processor, realizing a complete functional chain from video acquisition, skeleton extraction, region matching, differentiated risk calculation, two-level false alarm verification to time-series early warning and hierarchical alarm output. It has the advantages of high computing efficiency, flexible deployment, and strong real-time performance, and can be widely used in safety monitoring scenarios of rail transit stations.
[0057] It should be noted that the above embodiments are preferred embodiments of the present invention. The processor can be a single processor or a multi-processor cluster; the storage device can be an independent storage device or a distributed storage system; the program can be pre-stored in the storage device or downloaded and updated from a cloud server via a wired or wireless network. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.
[0058] Another embodiment of this application provides a computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire station surveillance video stream, extract the coordinates of key points of the human skeleton of passengers in the video frames, and generate skeleton data; Obtain a semantic parameter library for the station area, which includes the coordinate mapping relationship of each functional area, as well as the set of attitude feature weights and risk judgment thresholds corresponding to each functional area. Based on the passenger's location, a corresponding functional area is matched, and the posture feature weight set corresponding to that functional area is called to perform weighted calculation on the skeletal data to obtain an initial risk value. For passengers whose initial risk value exceeds the risk judgment threshold of the corresponding functional area, a false alarm check is performed. The false alarm check includes correcting the initial risk value based on the regional attributes of the functional area where the passenger is located, and / or performing behavioral reasoning based on the spatial positional relationship between objects in the passenger's surrounding scene and key points of the human skeleton to correct the risk value and output a valid risk value. Alarm information is generated and output based on the effective risk value.
[0059] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of this disclosure, and all such changes and modifications will fall within the scope of protection of this invention.
Claims
1. A method for early warning of passenger falls in rail transit stations, characterized in that, Includes the following steps: Acquire station surveillance video stream, extract the coordinates of key points of the human skeleton of passengers in the video frames, and generate skeleton data; Obtain a semantic parameter library for the station area, which includes the coordinate mapping relationship of each functional area, as well as the set of attitude feature weights and risk judgment thresholds corresponding to each functional area. Based on the passenger's location, a corresponding functional area is matched, and the skeletal data is weighted by calling the posture feature weight set corresponding to the functional area to obtain an initial risk value. For passengers whose initial risk value exceeds the risk judgment threshold of the corresponding functional area, a false alarm check is performed. The false alarm check includes correcting the initial risk value based on the regional attributes of the functional area where the passenger is located, and / or performing behavioral reasoning based on the spatial positional relationship between objects in the passenger's surrounding scene and key points of the human skeleton to correct the risk value and output a valid risk value. Alarm information is generated and output based on the effective risk value.
2. The passenger fall warning method for rail transit stations according to claim 1, characterized in that, The false alarm verification also includes: Level 1 verification: The initial risk value is attenuated and corrected based on the regional attributes of the functional area where the passenger is located; Level 2 verification: Identify scene objects within a preset range around the passenger, determine the passenger's behavior type based on the spatial relationship between key points of the human skeleton and the scene objects, and correct the risk value based on the behavior type. In the second level of verification: When the spatial relationship between the key points of the human buttocks and the seat meets the preset sitting judgment conditions, it is judged as a normal sitting behavior and the risk value is cleared to zero. When the spatial relationship between the key points of a person's hand and the object dropped on the ground meets the preset bending judgment conditions, it is judged as bending down to pick up the object, and the risk value is reduced by the first proportion. When the spatial relationship between the human body and the suitcase meets the preset sorting judgment conditions, it is judged as bending over to sort luggage behavior, and the risk value is reduced by the second proportion.
3. The passenger fall warning method for rail transit stations according to claim 1, characterized in that, Also includes: Obtain the skeletal temporal sequence of multiple consecutive frames; Extract gait abnormality precursor features from the skeletal time sequence; The gait abnormality precursor features are input into a temporal neural network for encoding, and a preliminary risk score is output. When the pre-risk score exceeds the warning threshold, a pre-warning signal is generated.
4. The passenger fall warning method for rail transit stations according to claim 3, characterized in that, The gait abnormality precursor features include at least one of the following: stride abrupt change rate, lateral sway amplitude of center of gravity, duration of trunk tilt, and degree of gait continuity interruption; Different functional areas are configured with different warning thresholds, and the warning threshold for high-risk functional areas is lower than that for low-risk functional areas.
5. The passenger fall warning method for rail transit stations according to claim 1, characterized in that, The functional areas are divided into high-risk areas, medium-risk areas and low-risk areas according to their security risk levels; The high-risk area is configured with a first weight set and a first judgment threshold. In the first weight set, the weight of the trunk tilt feature is greater than the weight of the limb extension feature, and the first judgment threshold is lower than the benchmark threshold, so that the high-risk area is more sensitive to abnormal trunk tilt. The medium-risk area is configured with a second weight set and a second judgment threshold, and the weights of each posture feature in the second weight set are evenly distributed. The low-risk area is configured with a third weight set and a third judgment threshold. In the third weight set, the weight of the limb extension feature is greater than the weight of the trunk tilt feature, and the third judgment threshold is higher than the benchmark threshold, so that the low-risk area has a lower false alarm rate.
6. The passenger fall warning method for rail transit stations according to claim 1, characterized in that, The generated and output alarm information includes: Risk levels are determined based on the magnitude of the effective risk value and / or the importance of the functional area where the passenger is located; Match the corresponding alarm output method according to the risk level; The system outputs tiered alarm information to the station operation system through a standardized interface.
7. A passenger fall warning system for rail transit stations, characterized in that, include: The video acquisition module is used to acquire the station's surveillance video stream; The human pose extraction module is used to extract the coordinates of key points of the passenger's human skeleton from the video stream and generate skeletal data; The region semantic labeling module is used to store the coordinate mapping relationship of each functional region, as well as the set of pose feature weights and risk judgment thresholds corresponding to each functional region; The risk calculation module is used to match the corresponding functional area based on the passenger's location, call the corresponding posture feature weight set to perform weighted calculation on the skeletal data, and obtain an initial risk value. The false alarm verification module is used to correct the risk value of passengers whose initial risk value exceeds the corresponding risk judgment threshold based on the regional attributes and / or the spatial positional relationship between scene objects and key points of the human skeleton, and output a valid risk value. The alarm output module is used to generate and output alarm information based on the effective risk value; The temporal risk warning module is used to acquire a continuous multi-frame skeletal temporal sequence, extract gait abnormality precursor features from the skeletal temporal sequence, input the gait abnormality precursor features into a temporal neural network for encoding, output a preliminary risk score, and generate a preliminary warning signal when the preliminary risk score exceeds a warning threshold.
8. The passenger fall warning system for rail transit stations according to claim 7, characterized in that, The alarm output module classifies risk levels based on the magnitude of the effective risk value and / or the importance of the functional area where the passenger is located, matches the corresponding alarm output method according to the risk level, and outputs graded alarm information to the station operation system through a standardized interface.
9. A passenger fall warning device for rail transit stations, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.
