Pedestrian tumble detection method and system based on attitude estimation

Abstract

The invention discloses a pedestrian falling detection method and system based on attitude estimation, and relates to the technical field of computer vision. The method comprises the following steps: initializing an attitude estimation model, and obtaining a video stream for frame-by-frame processing to obtain a current frame; performing human body detection on the current video frame based on the posture estimation model, judging whether a human body is detected, and if yes, obtaining a plurality of key points and coordinate information of the key points; when the number of the plurality of key points is greater than or equal to a preset number, triggering a deep tumble linkage analysis instruction; and pedestrian fall detection is carried out based on the deep fall linkage analysis instruction, and a fall early warning instruction is obtained. The technical problems of low precision, poor real-time performance and high false alarm rate of a fall detection method in the prior art due to factors such as illumination variation and shielding are solved, and the technical effect of high-robustness and low-delay accurate fall identification is achieved through high-precision sequential attitude analysis, dynamic threshold adjustment and multi-modal fusion.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method and system for pedestrian fall detection based on pose estimation. Background Technology

[0002] With the increasing aging of the global population, the health of the elderly has become a major social concern. Falls are one of the leading causes of injury and death among the elderly, especially those living alone or with limited mobility, who are at higher risk of falls. Traditional fall detection methods mostly rely on hardware sensors, such as accelerometers and gyroscopes, but these methods face problems such as insufficient accuracy, high false alarm rates, and poor real-time performance in practical applications. Furthermore, with the rapid development of intelligent monitoring and artificial intelligence technologies, fall detection methods based on computer vision are gradually gaining attention. Through pose estimation technology, it is possible to detect changes in pedestrian movement in real time and accurately from video streams to determine whether a fall has occurred, without relying on wearable devices or additional sensors.

[0003] Current computer vision technologies, especially deep learning-based pose estimation methods, have made significant progress. By accurately detecting key points on the human body, changes in posture can be analyzed, effectively identifying specific behaviors such as falls. However, existing technologies still face many challenges in practical applications. For example, detection accuracy is limited in complex environments such as those with changing lighting, occlusion, or multiple people, and real-time performance is insufficient, hindering rapid response to emergencies. Furthermore, existing methods typically lack effective temporal analysis mechanisms, making it difficult to distinguish falls from other similar behaviors (such as squatting or lying down), and they also struggle to provide complete video recording and cloud integration capabilities. This limits their widespread application in scenarios such as medical monitoring, elderly care, and public place security monitoring. Summary of the Invention

[0004] This application provides a pedestrian fall detection method and system based on pose estimation, which solves the technical problems of low accuracy, poor real-time performance and high false alarm rate in existing fall detection methods due to factors such as changes in lighting and occlusion.

[0005] The first aspect of this application provides a pedestrian fall detection method based on pose estimation, the method comprising:

[0006] Initialize the pose estimation model and acquire the video stream for frame-by-frame processing to obtain the current frame; perform human detection on the current video frame based on the pose estimation model to determine whether a human body is detected. If so, acquire multiple key points and their coordinate information; when the number of the multiple key points is greater than or equal to a preset number, trigger a deep fall linkage parsing instruction; perform pedestrian fall detection based on the deep fall linkage parsing instruction to obtain a fall warning instruction.

[0007] A second aspect of this application provides a pedestrian fall detection system based on pose estimation, the system comprising:

[0008] Frame-by-frame processing module: Initializes the pose estimation model and acquires the video stream for frame-by-frame processing to obtain the current frame; Human detection module: Performs human detection on the current video frame based on the pose estimation model, determines whether a human body is detected, and if so, acquires multiple key points and their coordinate information; Key point judgment module: Triggers a deep fall linkage parsing instruction when the number of the multiple key points is greater than or equal to a preset number; Pedestrian fall detection module: Performs pedestrian fall detection based on the deep fall linkage parsing instruction and obtains a fall warning instruction.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] First, the pose estimation model is initialized, and each frame of the acquired video stream is processed individually. In each frame, the model performs human detection, determining whether a human is detected. If detected, multiple key points of the human body and their coordinates are extracted. When the number of detected key points reaches a preset standard, a deep fall detection linkage analysis command is triggered. Based on this command, fall detection is performed, and a fall warning command is ultimately generated. The entire process, through a closed-loop design of initialization, detection, analysis, and warning, achieves end-to-end processing from video input to accurate fall behavior recognition, balancing detection accuracy, real-time performance, and environmental adaptability, making it suitable for various applications such as intelligent monitoring and elderly care. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic flowchart of a pedestrian fall detection method based on pose estimation provided in an embodiment of this application.

[0013] Figure 2 A schematic diagram of the pedestrian fall detection system based on posture estimation provided in this application embodiment.

[0014] Figure labeling: Frame-by-frame processing module 11, Human body detection module 12, Key point judgment module 13, Pedestrian fall detection module 14. Detailed Implementation

[0015] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0016] Example 1, as Figure 1 As shown, this application provides a pedestrian fall detection method based on pose estimation, the method including:

[0017] Initialize the pose estimation model and acquire the video stream for frame-by-frame processing to obtain the current frame.

[0018] In this embodiment, a pre-trained YOLOv11 pose estimation model is first initialized. This involves loading the model weights and network structure into memory or a dedicated computing device, such as a GPU, and configuring relevant inference parameters, such as input image size, confidence threshold, and non-maximum suppression parameters, to put the model into a ready state. This pose estimation model has been jointly trained and optimized using multi-source datasets, such as COCO, MPII, and a self-collected fall dataset, and possesses the ability to accurately detect 17 key points of the human body, such as head, torso, and limb nodes. Subsequently, a real-time video stream is acquired through a video capture unit. This video stream can originate from a surveillance camera, webcam, or local video file. The video stream is read frame by frame and decoded into RGB image frames. Each frame undergoes preprocessing operations, including resizing to meet model input requirements and normalizing pixel values. Then, each frame is processed sequentially in chronological order, and the image data of the current frame is extracted as input to the pose estimation model, ensuring continuous analysis of the video stream and real-time tracking of pedestrian poses, providing fundamental data support for subsequent key point detection and fall behavior analysis.

[0019] Furthermore, the pose estimation model is a YOLOv11 pose estimation model based on deep learning, which is used to simultaneously output the human target position and the corresponding human key point information; wherein, the YOLOv11 pose estimation model adopts a multi-dataset linkage training mechanism and uses K-fold cross-validation to evaluate the model performance.

[0020] Preferably, the pose estimation model used is the YOLOv11 model based on deep learning. This model integrates pose estimation tasks on the basis of the classic single-stage object detection framework, forming an end-to-end unified architecture. Its core process is that after receiving the input image, the YOLOv11 pose estimation model performs multi-level feature extraction through its improved backbone network, and then uses an enhanced Feature Pyramid Network (FPN) to fuse semantic information at different scales. Finally, the model outputs two types of information in parallel through an optimized composite detection head. One type is the bounding rectangle of the human target, including its position coordinates and presence confidence. These position coordinates are usually the center point coordinates, width, and height. The other type is the coordinates and confidence of a set of preset human key points corresponding to each detected human bounding box, such as the nose, left eye, and right eye among the head key points, the neck, left shoulder, and right shoulder among the torso key points, and the left hip, right hip, and left knee among the lower limb key points. This integrated design enables simultaneous human localization and pose estimation in a single forward inference, ensuring high efficiency and real-time performance. When training this YOLOv11 pose estimation model, multiple publicly available large-scale human pose datasets, such as COCO Keypoints, MPII Human Pose, and AI Challenger, are first integrated with a self-collected and labeled dataset of simulated fall scenarios. This multi-source data synergy combines the rich diversity of general scenarios with the specificity of the target scenario (fall), aiming to improve the model's generalization ability in different environments and the accuracy of fall pose recognition. Subsequently, the system trains the YOLOv11 pose estimation model in two stages.

[0021] In the first stage, pre-training is performed. This involves loading and preprocessing a large-scale open-source human pose dataset. For each input image, it is normalized and adjusted to a fixed size required by the model, such as 640x640 pixels. Basic geometric and photometric data augmentation strategies are applied, including random horizontal flipping, small-amplitude rotation and scaling, and fine-tuning of brightness and contrast, to initially improve the model's adaptability to changes in viewpoint and lighting. Next, the model initializes its weights with its default improved backbone network and enhanced feature pyramid network, and configures a composite detection head designed to simultaneously output the coordinates and confidence of the human bounding box, as well as the coordinates and visibility confidence of 17 keypoints for each instance. Training is performed end-to-end using a composite loss function that combines bounding box regression loss, target confidence loss, and keypoint coordinate regression loss. This composite loss function is a weighted mean squared error. AdamW or SGD with momentum is used as the optimizer, along with learning rate scheduling strategies such as cosine annealing, to iteratively update the model parameters over multiple epochs. By being fully trained on such large-scale, diverse general datasets, the model gradually mastered the ability to accurately locate human bodies and infer the spatial relationships of their standard key points in various common scenarios, laying a solid foundation for its efficient fine-tuning in specific fall detection tasks.

[0022] In the second phase, fine-tuning is performed. The model weights, acquired in the first phase and possessing general human perception capabilities, are used as high-starting-point initialization parameters. The training data is switched to a self-collected simulated fall dataset specifically designed to capture the entire posture process from standing imbalance to complete fall. Before training, these datasets undergo targeted enhancements, such as simulating motion blur, extreme perspectives, and random occlusion of key torso regions common during falls. This forces the model to learn to infer overall posture based on local information under non-ideal conditions. The core goal of fine-tuning is to make the model focus on learning abnormal posture patterns corresponding to fall behavior. This includes deviations in the spatial distribution of keypoints from the norm, abnormal distributions of keypoint confidence, and dramatic temporal dynamics of keypoint movement between adjacent frames. To address this, the training strategy is specifically adjusted: the optimizer's initial learning rate is set lower for finer parameter updates; the weights of the keypoint regression component in the loss function are appropriately increased, and weighted attention is introduced to fall keyframes; furthermore, a temporal negative sampling strategy is employed to mix non-fall "lying down" and "sitting down" sequences in batches to enhance the model's ability to distinguish similar static postures. Through multiple rounds of iterative fine-tuning on this self-collected simulated fall dataset, the model gradually transfers and specializes its general knowledge, ultimately becoming a dedicated posture estimation model capable of accurately identifying fall features and remaining robust to normal posture changes, providing accurate and reliable spatiotemporal keypoint information input for subsequent fall determination logic.

[0023] In the final stage of model training or during hyperparameter tuning, the entire training data is randomly divided into K subsets of roughly equal size, and K rounds of independent experiments are conducted based on these subsets. In each round, one subset is selected alternately as the validation set, and the remaining K-1 subsets are used as the training set. The model is then trained using the training set data from that round, and its performance metrics, such as keypoint detection accuracy (mAP) and model inference speed, are evaluated on the reserved validation set. After completing K rounds of validation, the K performance evaluation results are averaged to obtain the final performance evaluation report. This process makes full use of limited data and reduces evaluation bias caused by the randomness of a single data partition, thus providing a more reliable and objective estimate of the model's generalization ability and performance stability. Through the above multi-dataset linked training and rigorous K-fold cross-validation evaluation, the final pose estimation model possesses high accuracy, strong robustness, and good generalization ability, providing reliable human pose information input for subsequent fall detection analysis.

[0024] Human detection is performed on the current video frame based on the pose estimation model. If a human body is detected, multiple key points and their coordinate information are obtained.

[0025] In one embodiment, the current video frame is first input into an initialized pose estimation model. This model extracts image features through an improved backbone network and combines it with an enhanced feature pyramid to achieve multi-scale feature fusion, ensuring accurate detection of human targets. The model output includes the location of the human target and the corresponding keypoint coordinates. Subsequently, a threshold judgment is applied to the bounding box confidence. If the bounding box confidence is greater than or equal to the corresponding preset threshold, such as 0.5, it is determined that a human has been detected in the current frame; otherwise, it is considered that no human exists. When a human is detected, the model simultaneously outputs the coordinates of 17 keypoints corresponding to that human, each keypoint containing two-dimensional coordinates and a confidence value. To ensure the validity of keypoints, a keypoint quality verification is performed, filtering out keypoints with confidence values ​​less than the corresponding preset threshold, such as 0.3, and requiring at least 6 valid keypoints to be retained to ensure the integrity of pose information. In addition, relative coordinate calculation is used to unify the keypoint coordinates in images of different resolutions to a standard coordinate system, for example, normalizing to the range [-1,1] with the image center as the origin, and calculating the human height, i.e., the difference between the maximum and minimum Y values, to eliminate the influence of viewpoint changes. Finally, the output includes detection results containing parameters such as the target human body location, effective key point coordinates, and human body height, providing structured data support for subsequent fall status analysis.

[0026] Furthermore, if no human body is detected, a fall status recognition instruction is triggered; if a fall occurs, the fall time is detected; when the fall time is greater than or equal to a preset threshold, a fall warning instruction is triggered; wherein, the fall warning instruction includes triggering a fall alarm and an API alarm.

[0027] Preferably, a fall detection command is triggered when no human body is detected in the current video frame. Upon triggering, the state machine's historical records are retrieved to check if the preceding frame was in a fall state. If the preceding frame was in a normal state, subsequent steps are skipped; if the preceding frame was in a fall state, the fall time detection process begins. This involves extracting the fall state start timestamp from the state machine (e.g., the time point when the fall was first detected) and calculating the time difference with the current frame's timestamp to obtain the fall duration. When the fall duration is greater than or equal to a preset threshold (e.g., 10 seconds), it is considered a valid fall event; otherwise, it is considered a temporary state fluctuation and no warning is triggered. Once a valid fall event is identified, a fall warning command is triggered. This command uses hardware modules, such as buzzers and LED lights, to issue an audible and visual alarm, alerting personnel on site. The audible and visual alarm can be configured with intermittent beeping, flashing frequencies, etc., to differentiate between different risk levels. Simultaneously, this fall warning command also sends structured alert data to the mobile application or cloud platform via a RESTful API, including key information such as event time, location, video clip ID, and fall duration. API requests must include an authentication token to ensure data transmission security. If the fall duration does not reach the threshold, the fall state is maintained while monitoring continues; if a human body is detected again in subsequent frames and the posture returns to normal, the state machine is reset to the normal state, and temporary data is cleared. If the fall duration exceeds the threshold, after triggering the warning, the state machine is marked as processed to avoid repeated alarms. Simultaneously, event logging is initiated, recording the warning trigger time, processing results, and subsequent feedback in detail, forming a complete technical closed loop. This enables fall state recognition and warning triggering even when no human body is detected, ensuring robustness in complex scenarios.

[0028] When the number of the multiple key points is greater than or equal to the preset number, a deep fall linkage analysis command is triggered.

[0029] In one embodiment, after completing human detection and acquiring key point information for the current video frame, the number of acquired key points is statistically analyzed and judged. Specifically, the number of key points is first counted and compared with a pre-set threshold. When the number of valid key points is greater than or equal to the preset threshold, it indicates that the human posture information in the current video frame is highly complete and meets the requirements for complete human structural information in subsequent fall analysis. Subsequently, a deep fall linkage analysis command is triggered. This command initiates the subsequent fall behavior analysis process, including calculating human height and key point movement amplitude, and comparing human height. Conversely, if the number of valid key points is less than the preset threshold, it indicates that the current human posture information is incomplete or unreliable, insufficient to support accurate fall analysis. In this case, a fall status recognition command is triggered to determine whether a valid fall event exists. By only entering the deep analysis stage when the number of key points meets the preset condition, misjudgments caused by missing key points or incomplete posture information can be effectively avoided, thereby improving the stability and accuracy of the fall detection process and ensuring that subsequent fall detection logic can be executed based on reliable data.

[0030] Furthermore, when the number of the multiple key points is greater than or equal to a preset number, a deep fall-linked analysis instruction is triggered, including:

[0031] The human body height and the range of movement of key points are calculated based on the coordinates of multiple key points. When the human body height is less than or equal to a preset threshold, a deep fall linkage analysis command is triggered. When the human body height is greater than the preset threshold and an alarm has been triggered, a video continuous recording command is executed to record the safety status and obtain the next current frame. When the human body height is greater than the preset threshold and no alarm has been triggered, all states are reset, the safety status is recorded, and the next current frame is obtained.

[0032] Preferably, the head vertices, such as the nose and the bottom of the feet, are selected from the detected valid key points. The difference in their Y-coordinates is calculated as the human height. During the calculation, a relative coordinate transformation is used to map the original pixel coordinates to a standard coordinate system, eliminating the influence of camera viewpoint and resolution differences on the height calculation. For example, affine transformation is used to correct image distortion, or the reasonableness of the height is verified by prior knowledge of human proportions (e.g., the head accounts for 1 / 8 of the height). Then, the coordinates of the corresponding key points in the current frame and the previous frame are differentiated to calculate the movement range of the key points. Subsequently, a preset threshold is set based on the normal human height range observed over an initial period, generally 2 / 3 of the normal height. When the human height is less than or equal to the preset threshold, it is determined to be a sudden drop in height. At this time, a deep fall linkage analysis command is triggered, which activates the state machine to enter the fall analysis mode, that is, to confirm whether a fall event has occurred based on the movement range of the key points. When the human height is greater than the preset threshold and an alarm has been sent, a video continuous recording command is executed to record the safety status and save it to local or cloud storage, while continuing to process the next frame to monitor subsequent state changes. If the height of a person exceeds a preset threshold and no alarm is recorded, all state variables, such as the fall status flag and timer, are reset. The safe state is recorded, and the next frame is acquired for a new round of detection, avoiding misjudgments caused by residual states. In summary, by using height as the first-level rapid filtering condition, it achieves rapid filtering of normal states and efficient focusing on potential abnormal states. This provides precise triggering conditions for subsequent, more refined and resource-intensive deep fall-related analysis, balancing detection accuracy, real-time performance, and environmental adaptability.

[0033] Furthermore, when the height of a person is less than or equal to a preset threshold, a depth fall linkage analysis command is triggered, including:

[0034] When the height of the human body is less than or equal to a preset threshold, it is determined whether the movement range of the key point is less than the preset range threshold. If so, the height recognition result is a fall, and the deep fall linkage analysis instruction is to perform linkage analysis in the fall state; if not, the height recognition result is lying down, and the deep fall linkage analysis instruction is to perform linkage analysis in the lying down state.

[0035] Optionally, when the body height is less than or equal to a preset threshold, the overall posture of the body is determined to have changed from an upright state to a near-horizontal state. At this point, further behavior differentiation analysis is initiated to distinguish between falling behavior and similar posture behaviors such as voluntarily lying down. Specifically, the movement amplitude of key points on the body is compared with a preset amplitude threshold. When the movement amplitude of key points is less than or equal to the preset amplitude threshold, it indicates that after the body height rapidly decreases, the positional changes of key points tend to stabilize, which is consistent with the typical characteristic of the body remaining still for a short period of time after a fall. In this case, the height recognition result is determined to be a fall, and the deep fall linkage analysis instruction is set to perform linkage analysis in the fall state to trigger subsequent fall state confirmation, time duration judgment, and alarm processing procedures. When the movement amplitude of key points is greater than the preset amplitude threshold, it indicates that the body still has obvious continuous movement characteristics during the process of decreasing height, which usually corresponds to the behavior of the body voluntarily lying down, turning over, or adjusting posture. In this case, the height recognition result is determined to be lying down, and the deep fall linkage analysis instruction is set to perform linkage analysis in the lying down state for subsequent tracking and judgment of the lying down state, without directly triggering a fall alarm. By introducing a key point movement amplitude judgment mechanism, it is possible to effectively distinguish between falling and similar posture behaviors such as lying down, thereby significantly reducing the false alarm rate and improving the accuracy of fall detection.

[0036] Pedestrian fall detection is performed based on deep fall linkage analysis instructions to obtain fall warning instructions.

[0037] In one embodiment, after obtaining a deep fall-linked analysis command, the corresponding analysis logic is selected based on the command type. If the command is a fall-state linked analysis, it determines whether the human body is currently in a fall state, i.e., it checks whether a fall has been recorded in previous consecutive frames. If the human body was already in a fall state, the fall state marker will be maintained, and the fall duration timer will be updated to determine whether the fall duration has reached the alarm threshold. If the human body was not in a fall state, the fall state will be immediately set to true, the fall start time will be recorded, and the video recording function of the fall event will be started to ensure that the key frames before and after the fall are completely saved. If the deep fall-linked analysis command is a lying-down state, the lying-down behavior will be recorded, but a fall warning command will not be triggered immediately. Instead, the fall timer will be reset, the human body posture changes will be continuously monitored, and the state marker will be updated when the key points are stable or the human body returns to an upright position. The entire process ensures the continuity and real-time performance of the fall detection logic, while providing complete data support and event management through video recording and alarm mechanisms. Through the above process, it is possible to accurately distinguish between falls and lying-down behaviors, generate reliable fall warning commands, and provide an execution basis for real-time safety monitoring.

[0038] Furthermore, pedestrian fall detection is performed based on deep fall linkage analysis commands to obtain fall warning commands, including:

[0039] When the deep fall linkage parsing instruction is to perform linkage parsing in a fall state, it is determined whether the previous state was a fall state. If so, the fall state recognition instruction is triggered; otherwise, the fall state is set to true, and video recording is performed, the iteration start time is recorded, and the fall state is triggered.

[0040] Optionally, when the deep fall linkage analysis command is to perform linkage analysis under fall conditions, the fall status flag is read first. This fall status flag is used to persistently record whether the target human body has been identified as entering a potential fall state within the current monitoring period. If the fall status flag is true, it indicates that the human body has been confirmed to have entered a potential fall state in a previous video frame, and the corresponding monitoring timing and recording process has been initiated. At this time, the fall status recognition command will be triggered to perform the same fall time detection as mentioned above, thereby ensuring the continuity of status monitoring. If the fall status flag is false, it indicates that the current frame is the first time that the human body has been detected to meet the conditions for triggering deep fall analysis, namely, a significant decrease in height and a tendency for the posture to stabilize. This marks the beginning of a new potential fall event. At this time, the fall status flag will be set to true to lock and track this event. Subsequently, a dedicated event video recording process is immediately started or activated. This process usually saves the current frame and several previous frames together and begins to continuously record subsequent frames to generate a complete video evidence sequence from before the event to after the event. In addition, the system timestamp is recorded as the start time of this fall event iteration. This timestamp is the sole basis for calculating the fall duration and determining whether the alarm threshold has been reached. After the above initialization is completed, a fall status recognition command is triggered, transitioning to the subsequent continuous analysis and judgment phase. In summary, a persistent status flag can intelligently distinguish the start of a new fall event from the continuation of an existing fall event, avoiding repeated event initialization and false alarms, thereby achieving a high-precision, low-false-alarm fall detection status management mechanism.

[0041] Furthermore, this includes:

[0042] When the deep fall linkage parsing instruction is to perform linkage parsing in the lying state, it is determined whether the previous state was a fall state. If so, the lying state is recorded and the fall state recognition instruction is triggered; otherwise, the fall time is reset and the next current frame is obtained.

[0043] Optionally, when the deep fall-linked analysis instruction is to perform analysis in a lying-down state, a dedicated processing logic for lying-down behavior is initiated. Specifically, the stored fall state flag is first queried to confirm whether the human body has been marked as being in a fall monitoring state before the current judgment of lying down. When the fall state flag is true, it indicates that the human body may have undergone a process from suspected fall to actively adjusting to lying down. At this time, the posture of the current frame will be explicitly recorded as lying down, thus downgrading the potential, alarm-requiring continuous fall event to an active, potentially harmless low-posture behavior. At the same time, the fall state recognition instruction will still be triggered to continuously monitor and analyze the current lying-down state to prevent misjudgment of fall recovery ability. When the fall state flag is false, it indicates that the human body directly and actively enters the lying-down posture from normal activity without triggering the previous fall judgment conditions. At this time, all timers related to the fall event will be reset or cleared, interrupting any fall-oriented timing sequences that may have accumulated due to brief posture changes, preventing the erroneous continuation of historical data. Next, the system acquires and processes the next video frame, returning to the basic human detection and pose analysis loop to avoid unnecessary deep computation. In summary, through precise state tracking and conditional branching, it effectively distinguishes between lying down and voluntary lying down, which evolve from a fall. This not only prevents voluntary lying down from being misjudged as a fall and triggering alarms, but also maintains reasonable monitoring continuity during state transitions, ensuring the intelligence and reliability of the system's behavior.

[0044] In summary, the embodiments of this application have at least the following technical effects:

[0045] First, the pose estimation model is initialized, and the video stream is acquired and processed frame by frame to obtain the current frame. Next, based on the pose estimation model, human detection is performed on the current video frame to determine if a human is detected. If so, multiple key points and their coordinates are acquired. Then, when the number of these key points is greater than or equal to a preset number, a deep fall detection linkage command is triggered. Finally, pedestrian fall detection is performed based on the deep fall detection linkage command to obtain a fall warning command. This method solves the technical problems of low accuracy, poor real-time performance, and high false alarm rate in existing fall detection methods due to factors such as changes in lighting and occlusion. It achieves high robustness and low latency in accurate fall recognition through high-precision temporal pose analysis, dynamic threshold adjustment, and multimodal fusion.

[0046] Example 2, based on the same inventive concept as the pedestrian fall detection method based on pose estimation in the previous examples, such as... Figure 2 As shown, this application provides a pedestrian fall detection system based on pose estimation, the system comprising:

[0047] Frame-by-frame processing module 11: Initializes the pose estimation model and acquires the video stream for frame-by-frame processing to obtain the current frame; Human detection module 12: Performs human detection on the current video frame based on the pose estimation model, determines whether a human body is detected, and if so, acquires multiple key points and their coordinate information; Key point judgment module 13: Triggers a deep fall linkage parsing instruction when the number of the multiple key points is greater than or equal to a preset number; Pedestrian fall detection module 14: Performs pedestrian fall detection based on the deep fall linkage parsing instruction to obtain a fall warning instruction.

[0048] Furthermore, the frame-by-frame processing module 11 is used to perform the following method:

[0049] The pose estimation model is a YOLOv11 pose estimation model based on deep learning, which is used to simultaneously output the human target position and the corresponding human key point information; wherein, the YOLOv11 pose estimation model adopts a multi-dataset linkage training mechanism and uses K-fold cross-validation to evaluate the model performance.

[0050] Furthermore, the human body detection module 12 is used to perform the following methods:

[0051] If no human body is detected, a fall detection command is triggered. If a fall occurs, the fall time is detected. If the fall time is greater than or equal to a preset threshold, a fall warning command is triggered. The fall warning command includes triggering a fall alarm and an API alarm.

[0052] Furthermore, the key point determination module 13 is used to perform the following method:

[0053] The human body height and the range of movement of key points are calculated based on the coordinates of multiple key points. When the human body height is less than or equal to a preset threshold, a deep fall linkage analysis command is triggered. When the human body height is greater than the preset threshold and an alarm has been triggered, a video continuous recording command is executed to record the safety status and obtain the next current frame. When the human body height is greater than the preset threshold and no alarm has been triggered, all states are reset, the safety status is recorded, and the next current frame is obtained.

[0054] Furthermore, the key point determination module 13 is used to perform the following method:

[0055] When the height of the human body is less than or equal to a preset threshold, it is determined whether the movement range of the key point is less than the preset range threshold. If so, the height recognition result is a fall, and the deep fall linkage analysis instruction is to perform linkage analysis in the fall state; if not, the height recognition result is lying down, and the deep fall linkage analysis instruction is to perform linkage analysis in the lying down state.

[0056] Furthermore, the key point determination module 13 is used to perform the following method:

[0057] When the deep fall linkage parsing instruction is to perform linkage parsing in the lying state, it is determined whether the previous state was a fall state. If so, the lying state is recorded and the fall state recognition instruction is triggered; otherwise, the fall time is reset and the next current frame is obtained.

[0058] Furthermore, the pedestrian fall detection module 14 is used to perform the following method:

[0059] When the deep fall linkage parsing instruction is to perform linkage parsing in a fall state, it is determined whether the previous state was a fall state. If so, the fall state recognition instruction is triggered; otherwise, the fall state is set to true, and video recording is performed, the iteration start time is recorded, and the fall state is triggered.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A pedestrian fall detection method based on pose estimation, characterized in that, The method includes: Initialize the pose estimation model and acquire the video stream for frame-by-frame processing to obtain the current frame; Human detection is performed on the current video frame based on the pose estimation model. If a human body is detected, multiple key points and their coordinate information are obtained. When the number of the multiple key points is greater than or equal to the preset number, a deep fall linkage analysis command is triggered; Pedestrian fall detection is performed based on deep fall linkage analysis instructions to obtain fall warning instructions.

2. The pedestrian fall detection method based on pose estimation as described in claim 1, characterized in that, The pose estimation model is a YOLOv11 pose estimation model based on deep learning, which is used to simultaneously output the human target position and the corresponding human key point information. The YOLOv11 pose estimation model employs a multi-dataset training mechanism and utilizes K-fold cross-validation to evaluate model performance.

3. The pedestrian fall detection method based on pose estimation as described in claim 1, characterized in that, If no human body is detected, a fall detection command is triggered. If a fall occurs, the fall time is detected. If the fall time is greater than or equal to a preset threshold, a fall warning command is triggered. The fall warning command includes triggering a fall alarm and an API alarm.

4. The pedestrian fall detection method based on pose estimation as described in claim 1, characterized in that, When the number of the multiple key points is greater than or equal to a preset number, a deep fall-linked analysis instruction is triggered, including: Calculate human height and key point movement range based on the coordinates of multiple key points; When the height of a person is less than or equal to a preset threshold, a depth fall linkage analysis command is triggered. When the height of a person exceeds a preset threshold and an alarm has been triggered, execute the instruction to continuously record video, record the safety status, and obtain the next current frame; When the height of a human body exceeds a preset threshold and no alarm is triggered, all states are reset, the safety status is recorded, and the next current frame is obtained.

5. The pedestrian fall detection method based on pose estimation as described in claim 4, characterized in that, When the height of the human body is less than or equal to a preset threshold, a depth fall linkage analysis command is triggered, including: When the height of the human body is less than or equal to the preset threshold, it is determined whether the movement range of the key point is less than the preset range threshold. If so, the height recognition result is a fall, and the depth fall linkage analysis instruction is to perform linkage analysis in the fall state. If not, the height recognition result is "lying down," and the depth fall linkage analysis instruction is to perform linkage analysis in the lying down state.

6. The pedestrian fall detection method based on pose estimation as described in claim 5, characterized in that, Pedestrian fall detection is performed based on deep fall linkage parsing instructions to obtain fall warning instructions, including: When the deep fall linkage parsing instruction is to perform linkage parsing in the fall state, it is determined whether the previous state was a fall state. If so, the fall state recognition instruction is triggered. If not, set the fall status to true, record video, record the iteration start time and fall status, and trigger the fall status recognition command.

7. The pedestrian fall detection method based on pose estimation as described in claim 5, characterized in that, include: When the deep fall linkage parsing instruction is to perform linkage parsing in the lying state, it is determined whether the previous state was a fall state. If so, the lying state is recorded and the fall state recognition instruction is triggered. If not, reset the fall time and get the next current frame.

8. A pedestrian fall detection system based on pose estimation, characterized in that, For implementing the pedestrian fall detection method based on pose estimation as described in any one of claims 1-7, the system comprises: Frame-by-frame processing module: Initializes the pose estimation model, acquires the video stream for frame-by-frame processing, and obtains the current frame; Human detection module: Based on the pose estimation model, human detection is performed on the current video frame to determine whether a human body is detected. If so, multiple key points and their coordinate information are obtained. Key point judgment module: When the number of the multiple key points is greater than or equal to the preset number, a deep fall linkage analysis command is triggered; Pedestrian fall detection module: Based on the deep fall linkage parsing command, pedestrian fall detection is performed to obtain fall warning command.

Images