Marathon cross-country event tumble detection method based on unmanned aerial vehicle and attitude estimation

Abstract

The invention provides a marathon cross-country event tumble detection method based on an unmanned aerial vehicle and attitude estimation, relates to the technical field of image recognition, constructs a P2 high-resolution detection branch, retains the highest-frequency spatial texture details in a human body image, and solves the problem of missing detection caused by feature aliasing of minimum targets such as wrists and ankles. According to the method, a C2PSA module is embedded in the deep layer of a backbone network, attention weights are concentrated in a human body target area, a distributed sensing coordinate classification architecture is introduced to process X-axis and Y-axis feature vectors, and the defect that correlation between coordinate axes is ignored in an existing method is overcome. According to the method, a parallel Aux Head auxiliary branch and a KLD loss function are introduced, so that the model can tolerate labeling errors in the optimization process. According to the method, false alarm filtering is carried out in combination with head motion logic, active stretching / shoelace tying and accidental falling are effectively distinguished, and the false alarm rate in practical application is greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a fall detection method for marathon cross-country races based on unmanned aerial vehicles and attitude estimation. Background Technology

[0002] With the deepening implementation of the national strategy for fitness for all, long-distance endurance events such as marathons and trail running have seen explosive growth worldwide. During these races, athletes frequently experience sudden situations such as falls, hypothermia, or even cardiac arrest due to exhaustion, heatstroke, or slippery surfaces, which can easily lead to serious safety accidents.

[0003] Currently, safety supervision for marathon and trail running events primarily relies on the traditional model of "manual patrols + fixed-point monitoring." Ground support mainly depends on volunteers, medical runners (first aid runners), and visual observation from fixed medical stations along the route. In recent years, some high-profile events have begun to introduce drones for aerial patrols. However, current drone applications are mainly limited to live aerial photography or manual visual inspections by pilots via image transmission screens, lacking automated intelligent analysis methods. This heavily reliant on manual labor makes it difficult to achieve full coverage when facing a wide racecourse, and manual monitoring is prone to fatigue and oversights, leading to delayed responses.

[0004] To address the efficiency issues of manual monitoring, computer vision-based intelligent analysis technologies have gradually become a research hotspot in the industry. Among these, human pose estimation is a core technical approach. Due to its extremely fast inference speed and small model size, YOLOv11-Pose has become the preferred base model for real-time motion analysis on edge devices. However, existing general-purpose algorithms often face serious technical bottlenecks when directly transferred to marathon scenarios. These bottlenecks manifest in the following four aspects: Loss of key features for extremely small targets: In marathon trail running events, the pixel ratio of participants in the video footage of monitoring equipment is extremely low, and the existing YOLOv11-Pose model cannot perceive the position of the limbs during inference.

[0005] Insufficient feature extraction capability in complex dynamic backgrounds: When faced with the complex background of a trail running track, existing models lack an effective attention focusing mechanism, making it difficult to separate the semantic features of the human torso and limbs from the surrounding similar environment.

[0006] Poor robustness of localization in occluded and blurred scenes: Existing direct coordinate regression methods lack the ability to express uncertainty when occlusion or blurring leads to unclear features, and often give a highly biased erroneous coordinate.

[0007] The loss function lacks physical constraints and spatial awareness: For "difficult samples" such as wrists and ankles, which are crucial for fall detection, the existing OKS Loss loss function lacks a targeted weighted optimization strategy.

[0008] In summary, although drone hardware and attitude estimation algorithms have both matured, there is currently a lack of a mature fall detection solution that can deeply integrate the two and is specifically designed for environments such as marathon trail running, which are characterized by "wide outdoor areas, complex backgrounds, long distances / high-altitude perspectives, and extremely small targets". Summary of the Invention

[0009] To address the shortcomings of existing technologies, the present invention aims to propose a fall detection method for marathon trail running races based on unmanned aerial vehicles (UAVs) and attitude estimation, comprising: Images of marathon trail running events are acquired and processed to obtain the first image; The first image is input into a pre-trained improved YOLOv11 model to obtain the predicted coordinates of the human joints of the runners in the marathon cross-country race. The improved YOLOv11 model introduces a spatial attention mechanism into the YOLOv11 model, adds a detection head and a DCC-based distribution-aware coordinate classification and prediction architecture. Based on the predicted coordinates of the human joints, calculate the trunk tilt angle θ, the normalized height of the center of gravity Hnorm, and the aspect ratio R of the circumscribed rectangle; Based on the torso tilt angle θ, the normalized height of the center of gravity, and the aspect ratio R of the circumscribed rectangle, determine whether a runner in a marathon cross-country race image has fallen.

[0010] Optionally, images of marathon trail running events are acquired and processed to obtain a first image, including: High-definition network cameras are deployed in accident-prone areas of marathon trail running events to capture images. In blind spots that cannot be covered by fixed equipment, industrial-grade drones equipped with three-axis gimbal cameras are used to capture images. An adaptive filling strategy is used to scale the captured images to obtain the first image.

[0011] Optionally, the first image is input into the improved YOLOv11 model to obtain the predicted coordinates of the human joints of athletes participating in a marathon trail running event, including: The first image is downsampled through the STEM layer to obtain the first feature map. The first feature map is downsampled to obtain the second feature map. The second feature map is downsampled to obtain the third feature map. The third feature map is downsampled to obtain the downsampled third feature map. The downsampled third feature map is processed through a spatial attention mechanism to obtain the fourth feature map. The fourth feature map is downsampled to obtain the downsampled fourth feature map. The downsampled fourth feature map is processed through a spatial attention mechanism to obtain the fifth feature map. The fifth feature map is upsampled to obtain the upsampled fifth feature map. The upsampled fifth feature map is then fused with the fourth feature map to obtain the first fused feature map. The first fused feature map is processed through a DCC-based distribution-aware coordinate classification and prediction architecture to obtain the predicted coordinates of human joints. Specifically, the first fused feature map is processed through a convolutional layer to obtain the X-axis feature vector Vx and the Y-axis feature vector Vy. The correlation matrix between Vx and Vy is calculated, and the correlation matrix is ​​input into the classifier to obtain the probability distribution vector Px on the X-axis and the probability distribution vector Py on the Y-axis. The weighted average of the probability distribution vectors Px on the X-axis and Py on the Y-axis is calculated to obtain the predicted coordinates of human joints.

[0012] Optionally, the convolutional kernel weights in the improved YOLOv11 model are iteratively updated using the total loss, which is calculated as follows: The first fusion feature map is upsampled to obtain the upsampled first fusion feature map. The upsampled first fusion feature map is then fused with the third feature map to obtain the second fusion feature map. The second fusion feature map is then upsampled to obtain the upsampled second fusion feature map. The upsampled second fusion feature map is then fused with the second feature map to obtain the third fusion feature map. The third fused image is passed through the detection head to obtain the first detection feature map; the first fused feature map is processed through the parallel lightweight auxiliary branch AuxHead to obtain the visibility result, which represents whether the position in the fourth feature map is visible; The visibility loss is calculated based on the joint optimized location loss function and the visibility results. Based on the true coordinates of the key points and the preset variance σ, a Gaussian distribution is generated as a soft label for the predicted coordinates. The KL divergence between the model's predicted coordinates and the soft label is calculated to obtain the KL divergence loss. The visibility loss and the KL divergence loss together constitute the total loss.

[0013] Optionally, based on the predicted coordinates of the human joints, calculate the trunk tilt angle θ, the normalized height of the center of gravity Hnorm, and the aspect ratio R of the circumscribed rectangle, including: In the predicted coordinates of human joints, connect the key points of the neck and the mid-hip point, calculate the angle between the connected vector and the vertical line of gravity to obtain the torso tilt angle θ; obtain the Y coordinate of the mid-hip point, calculate the relative height of the Y coordinate of the mid-hip point from the ground, divide the relative height by the height of the human detection box to obtain the normalized height of the center of gravity Hnorm; calculate the ratio of the width W to the height H of the human detection box to obtain the aspect ratio R of the bounding rectangle.

[0014] Optionally, based on the torso tilt angle θ, the normalized height of the center of gravity, and the aspect ratio R of the circumscribed rectangle, determine whether the runner in the image of a marathon trail running event has fallen, including: When θ > 45° and the normalized height of the center of gravity Hnorm decreases by more than a preset threshold ΔH within the time window, the runners in the image representing the marathon trail running event are in an unstable state. When an athlete in a marathon trail running image is in an unstable state and meets any of the target conditions, it indicates that the athlete in the marathon trail running image is in a fall state. When an athlete in a marathon cross-country race image is in a fallen state, if the head key point in the predicted coordinates of the human joints exhibits vertical reciprocating motion with an amplitude exceeding a preset threshold ΔH within a preset time period, the athlete's state in the marathon cross-country race image is reset to a normal state. If the average optical flow velocity of all key points in the predicted coordinates of the human joints is lower than the static threshold, and the fallen state is maintained for more than the safety confirmation time, the athlete in the marathon cross-country race image is confirmed to have fallen.

[0015] Optionally, the target conditions include: the aspect ratio R of the circumscribed rectangle is greater than 1.2; the torso tilt angle θ is greater than 65° and the duration exceeds a preset time.

[0016] The beneficial effects of adopting the above technical solution are as follows: This invention innovatively constructs a feature fusion path that runs directly from the P2 layer (1 / 4 downsampled) of the backbone network to the head. This design preserves the most frequent spatial texture details in human images, fundamentally solving the problem of missed detection of "extremely small targets" such as wrists and ankles due to feature aliasing in high-altitude drone-view or long-distance monitoring scenarios from the perspective of physical resolution, and significantly improving the recall rate of tiny key points.

[0017] This invention embeds a C2PSA module in the deep layers (Stage 4 / 5) of the backbone network. This module utilizes a spatial attention mechanism to force the network to focus attention weights on the human target region, actively suppressing background noise from the surrounding environment. This improvement enables the model to extract high-purity human semantic features even in dynamically changing aerial photography backgrounds. This invention abandons the traditional direct coordinate regression method and proposes a Distribution-Aware Coordinate Classification (DCC) architecture. A Cross-Attention mechanism is introduced to process the X-axis and Y-axis feature vectors, addressing the deficiency of existing methods such as SimCC in ignoring the correlation between coordinate axes. The localization task is transformed into a probability distribution classification task on discrete intervals. When a fall causes limb occlusion or blurring, the model can output a "fuzzy correct" result that conforms to the probability distribution, rather than a single "precise incorrect" coordinate, greatly improving the robustness of the algorithm.

[0018] In terms of training strategy, this invention introduces a parallel Aux Head auxiliary branch and a KLD loss function. Multi-task learning forces the model to explicitly learn the "visible / occluded" states of keypoints, and Gaussian soft labels are used instead of hard labels, allowing the model to tolerate labeling errors during optimization. Weighted optimization is also implemented for challenging samples such as wrists and ankles. This is combined with biomechanical features and a temporal state machine.

[0019] This invention comprehensively utilizes torso tilt angle, normalized height drop rate of center of gravity, and aspect ratio characteristics, combined with head movement logic for false alarm filtering. This logic design effectively distinguishes between "active stretching / tying shoelaces" and "accidental falls," significantly reducing the false alarm rate in practical applications. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the fall detection method for marathon trail running events based on unmanned aerial vehicles and attitude estimation in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the improved YOLOv11 model in an embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] To address the problems existing in the prior art, this invention provides a fall detection method for marathon trail running events based on drones and attitude estimation. This method aims to solve the problem of loss of features of "extremely small human targets" caused by long-distance monitoring and the high-altitude perspective of drones in wide-area outdoor scenarios, as well as the problem of inaccurate positioning caused by complex backgrounds (forests, crowds) and limb occlusion.

[0023] The primary objective of this invention is to provide a fall detection method for marathon trail running races based on UAV collaboration and improved attitude estimation. This method aims to overcome the shortcomings of existing general attitude estimation techniques in wide-area outdoor scenes, UAV aerial perspectives, and complex dynamic backgrounds, such as low detection rate of extremely small human targets, inaccurate key point localization, and weak anti-occlusion capabilities. Specifically, this invention strives to achieve the following specific technical objectives: First, it addresses the issue of feature loss for "extremely small targets" from high-altitude and long-distance perspectives. Addressing the extremely low pixel count of participants due to drone patrols or remote monitoring, this invention aims to construct a high-resolution detection branch directly connecting the backbone network's P2 layer to the head. This preserves high-frequency spatial details of human extremities (wrists, ankles) at the physical resolution level, fundamentally resolving feature aliasing and missed detection issues caused by excessively high downsampling rates, and significantly improving the recall rate of tiny targets.

[0024] Secondly, this invention addresses the challenges of background interference and feature extraction in complex outdoor environments. To mitigate dynamic background noise caused by tree obstructions, crowd clutter, and drone flight vibrations in off-road tracks, this invention aims to introduce a C2PSA (Cross-Stage Partial with Spatial Attention) spatial attention mechanism into the deep layers of a lightweight network. This forces the model to transform passively receiving features into actively focusing on the semantic features of the human torso and limbs, effectively suppressing environmental background interference and ensuring the purity of feature extraction.

[0025] Third, this invention addresses the localization failure caused by occlusion and blurring during a fall. To address the issues of violent limb folding, motion blur, and self-occlusion during a fall, this invention aims to abandon traditional coordinate regression methods and instead adopt a Distribution-Aware Coordinate Classification (DCC) architecture, combined with a cross-attention mechanism to handle coordinate correlation. Simultaneously, by incorporating the KLD soft-label loss function and an auxiliary supervision branch, the model can understand the physical characteristics of "occlusion," outputting reasonable predictions that conform to a probability distribution even when features are unclear, rather than simply providing erroneous coordinates, thereby significantly improving the robustness of localization.

[0026] Fourth, construct a ground-air integrated, high-real-time precision fall detection system. This invention aims to integrate multi-source data from drones and fixed locations, combining biomechanical characteristics (center of gravity, tilt angle) and temporal finite state machines to construct a fall logic algorithm that can effectively filter false falls (such as tying shoelaces, stretching), achieving millisecond-level response and alarm for safety hazards across the entire event area while ensuring the real-time operation of edge devices.

[0027] The method of this invention mainly includes six core steps: multi-source stereo image acquisition and preprocessing, C2PSA-based feature extraction, P2 Head-based multi-scale feature fusion, DCC-based distributed sensing coordinate prediction, KLD and AuxHead-based model training, and fall logic determination based on multi-dimensional spatiotemporal features. Combined with... Figure 1 This may include the following steps: Step 1: Acquire images of the marathon trail running event, process the images to obtain the first image; (i.e., multi-source stereo image acquisition and preprocessing). High-definition network cameras are deployed in high-accident areas of marathon trail running, such as sharp bends, steep slopes, and aid stations. The cameras are mounted at a height of 3-5 meters to provide fixed overhead or eye-level monitoring images. Images are captured by high-definition network cameras. In blind spots such as forests and canyons where fixed equipment cannot cover, industrial-grade drones equipped with three-axis gimbal cameras are used to capture images. The drones typically fly at an altitude of 20-50 meters to provide vertical overhead or large-angle oblique views.

[0028] An adaptive padding strategy is employed to scale the acquired images, uniformly scaling input images of different resolutions (e.g., 1080P, 4K) to the model input size (640×640) to obtain the first image. Specifically, to maintain the aspect ratio of the human body without distortion, an adaptive padding strategy is used, which scales the longer side to the target size proportionally and fills the shorter side with gray pixel values ​​(RGB: 114, 114, 114). The pixel values ​​are normalized from [0, 255] to the [0, 1] interval to accelerate network convergence.

[0029] Step 2: Input the first image into the improved YOLOv11 model to obtain the predicted coordinates of the human joints of the runners in the marathon cross-country race. The improved YOLOv11 model introduces a spatial attention mechanism into the YOLOv11 model, adds a detection head and a DCC-based distribution-aware coordinate classification and prediction architecture. Step 2 is feature extraction (introducing the C2PSA spatial attention mechanism). To accurately extract human features in complex off-road backgrounds (such as trees, grass, and shadows), this embodiment embeds C2PSA modules at the ends of Stage 4 (16x downsampling) and Stage 5 (32x downsampling) of the backbone network. The C2PSA module forces the network to ignore background noise and focus only on the human torso and limbs.

[0030] The processed 640x640 pixel first image is fed into the improved YOLOv11 backbone network. After initial downsampling in the stem layer to reduce computation, a 320×320 pixel image is output. This downsampling is then repeated through a series of layers (Stage 2 to Stage 5), ultimately outputting "feature maps" of different scales (P2, P3, P4, P5). Specifically, combined with... Figure 2 The first image is downsampled through the stem layer to obtain the first feature map. The first feature map is downsampled to obtain the second feature map. The second feature map is downsampled to obtain the third feature map. The third feature map is downsampled to obtain the downsampled third feature map. The downsampled third feature map is processed through a spatial attention mechanism to obtain the fourth feature map. The fourth feature map is downsampled to obtain the downsampled fourth feature map. The downsampled fourth feature map is processed through a spatial attention mechanism to obtain the fifth feature map. The fifth feature map is upsampled to obtain the upsampled fifth feature map. The upsampled fifth feature map is then fused with the fourth feature map to obtain the first fused feature map. The first fused feature map is upsampled to obtain the first fused feature map. The upsampled first fused feature map is then fused with the third feature map to obtain the second fused feature map. The second fused feature map is upsampled to obtain the second fused feature map. The upsampled second fused feature map is then fused with the second feature map to obtain the third fused feature map. Although the third fused feature map has weaker semantic information, it retains extremely high-frequency spatial texture details, which is crucial for locating wrists and ankles that are only a few pixels wide.

[0031] Multi-scale feature fusion and perception of extremely small targets based on P2 Head. Addressing the issue of extremely low pixel counts at the extremities (wrist, ankle) of the human body from a high-altitude UAV perspective, this embodiment significantly improves the feature fusion network and detection head, constructing a high-resolution detection branch directly connecting the P2 layer to the head. Traditional YOLOv11-Pose typically uses a PANet (Path Aggregation Network) structure to fuse features from three layers: P3 (1 / 8), P4 (1 / 16), and P5 (1 / 32). This invention adds a high-resolution fusion path of P2 (1 / 4 downsampling).

[0032] For the P2, P3, P4, and P5 feature maps, four detection heads at different scales are constructed. A new P2 head is added specifically to detect small targets (such as distant human figures) and fine-grained key points on the human body. P3-P5 heads are responsible for detecting medium- and large-scale targets. This design addresses the problem of small target features being "submerged" or aliased in deep networks at the physical resolution level, significantly improving the recall rate for extremities.

[0033] Specifically, the third fused image is passed through the first-scale detection head to obtain the first detection feature map; the second fused feature map is passed through the second-scale detection head to obtain the second detection map; and the fifth feature map is passed through the third-scale detection head to obtain the third detection map; wherein the first scale, the second scale, and the third scale are all different. Next, we will discuss distributed sensing coordinate classification prediction based on DCC. To address the "positioning drift" problem caused by limb occlusion and folding during a fall, this embodiment abandons the traditional direct coordinate regression scheme and introduces a DCC architecture as the pose estimation head. The DCC architecture models the keypoint localization task as a "classification task on a discrete grid".

[0034] Specifically, the 2D position (x, y) obtained from NeckP4 is decoupled into two independent 1D vector prediction tasks. The first fused feature map is processed by a DCC-based distribution-aware coordinate classification and prediction architecture to obtain the predicted coordinates of the human joints. Specifically, the first fused feature map is passed through a convolutional layer to obtain the X-axis feature vector Vx∈RW×K and the Y-axis feature vector Vy∈RH×K. The correlation matrix between Vx and Vy is calculated, allowing the X-axis prediction to reference the Y-axis feature distribution, and vice versa. For example, when predicting the x-coordinate of the "right wrist," the network considers whether its y-coordinate is in the "raised" or "drooping" region, thus eliminating unreasonable predictions.

[0035] The correlation matrix is ​​input into the classifier to obtain the probability distribution vector Px on the X-axis and the probability distribution vector Py on the Y-axis. The final coordinate values ​​are not taken from the single grid point with the highest probability, but are obtained through expectation calculation, that is, by calculating the weighted average of the probability distribution vector Px on the X-axis and the probability distribution vector Py on the Y-axis to obtain the predicted coordinates of the human joints. When occlusion occurs, the probability distribution will be "multi-peaked" or "flat". The model will tend to output the center of the region with the highest probability, rather than jumping directly to 0 or flying out of the image boundary like the regression method, which has extremely strong noise resistance.

[0036] During the training phase, in addition to the main detection head, a parallel lightweight auxiliary branch, AuxHead, is added. The first fused feature map is processed through the parallel lightweight auxiliary branch AuxHead to obtain the visibility result, which characterizes whether a position in the fourth feature map is visible. The visibility loss is calculated based on the jointly optimized location loss function and the visibility results, forcing the features extracted by the backbone network to include more semantic information about the "occlusion state," thereby feeding back into the localization task of the main branch. The Aux Head is only enabled during the training phase and removed during the inference phase through reparameterization or direct pruning, without increasing the inference time.

[0037] To address the issue of standard one-hot labels being too "absolute," this embodiment employs a loss function based on KL divergence. Based on the true coordinates of keypoints and a preset variance σ, Gaussian distributions are generated as soft labels for predicted coordinates, reflecting the inherent labeling errors and ambiguities in the keypoint locations. The KL divergence between the model's predicted coordinates and the soft labels is calculated, yielding the KL divergence loss. The visibility loss and KL divergence loss together constitute the total loss. Based on this total loss, the convolutional kernel weights of the improved YOLOv11 model are updated, introducing an OKs-aware Weighting mechanism. Higher loss weights (compared to all convolutional kernel weights in the entire network) are assigned to "difficult points" such as wrists and ankles, which are prone to detection errors; lower weights are assigned to easier points such as shoulders and hips. This achieves "targeted" optimization for difficult samples.

[0038] Step 3: Calculate the trunk tilt angle θ, the normalized height of the center of gravity Hnorm, and the aspect ratio R of the circumscribed rectangle based on the predicted coordinates of the human joints; In the predicted coordinates of human joints, connect the key points of the neck and the mid-hip point, calculate the angle between the connected vector and the vertical line of gravity to obtain the trunk tilt angle θ, which is less than 20° when standing or running normally; obtain the Y coordinate of the mid-hip point, calculate the relative height of the Y coordinate of the mid-hip point from the ground, divide the relative height by the height of the human detection box to obtain the normalized height of the center of gravity Hnorm; calculate the ratio of the width W to the height H of the human detection box to obtain the aspect ratio R of the bounding rectangle. Step 4: Determine whether the runner in the image of the marathon trail running event has fallen based on the torso tilt angle θ, the normalized height of the center of gravity, and the aspect ratio R of the circumscribed rectangle. When θ > 45° (significant torso tilt) and the normalized height of the center of gravity Hnorm decreases by more than a preset threshold ΔH (e.g., a 40% decrease) within a time window of Twin = 0.5s, the athlete in the image representing the marathon trail running event is in an unstable state. The P2 Head is used here to accurately locate the ankle, ensuring the accuracy of the ground position estimation, thereby accurately calculating the relative descent of the center of gravity and effectively eliminating the "slow bending over" movement.

[0039] When an athlete in a marathon trail running image is in an unstable state and meets any of the target conditions, it indicates that the athlete in the marathon trail running image is in a fall state. The target conditions include: the aspect ratio R of the circumscribed rectangle is greater than 1.2 (the human body is laid flat on the ground); the torso tilt angle θ is greater than 65° and the duration exceeds the preset duration Thold=1.0s; When a runner in a marathon cross-country race image is in a fallen state, if the head key points in the predicted coordinates of the human joints exhibit vertical reciprocating motion with an amplitude exceeding a preset threshold ΔH within a preset time period (feature matching: head undulation when the runner is tying his shoelaces), the runner's state in the marathon cross-country race image is reset to a normal state. If the average optical flow velocity (or inter-frame displacement velocity) of all key points in the predicted coordinates of the human joints is lower than the static threshold, and the runner remains in a fallen state for more than the safety confirmation time, the runner in the marathon cross-country race image is determined to have fallen. Once a fall is confirmed, the system immediately executes the following steps: It locks the target's Track ID and extracts short video clips from the video stream, including 5 seconds before and after the fall. Using the drone's GPS information or fixed camera calibration information, it calculates the geographical coordinates of the fall point. The alarm information (time, location, confidence level, and video clips) is then pushed via 4G / 5G network to the race command center's large screen and the nearest medical runner's handheld terminal.

[0040] This embodiment achieves the following technical breakthroughs by improving the network structure and introducing multi-dimensional logic judgment: P2 Head enables the drone to accurately capture the athlete's wrist and ankle at an altitude of 50 meters; C2PSA enables the algorithm to lock onto the human body in the shade of trees without being disturbed by the environment; DCC ensures that the coordinate prediction remains smooth and accurate when the athlete falls, rolls, or has limb obstruction; and the temporal logic effectively filters out false positive actions commonly seen on the field, such as tying shoelaces.

[0041] The key point of this invention is: 1. A high-resolution P2 detection pathway for UAVs / long-range perspectives was constructed; Unlike the conventional approach of the YOLOv11 model, which only utilizes features from layers P3 (1 / 8 downsampled) to P5 for detection, this invention innovatively constructs a feature fusion path that runs directly from the backbone network layer P2 (1 / 4 downsampled) to the head. This design preserves the most frequent spatial texture details in human images, fundamentally solving the problem of missed detection of "extremely small targets" such as wrists and ankles due to feature aliasing in aerial drone-based or long-distance monitoring scenarios, thus significantly improving the recall rate of tiny keypoints.

[0042] 2. Backbone network enhancement with embedded C2PSA spatial attention mechanism; To address the complex background interference caused by trees, grass, and people on marathon trail running courses, this invention embeds a C2PSA module in the deep layers (Stage 4 / 5) of the backbone network. This module utilizes a spatial attention mechanism to force the network to focus its attention weights on the human target area, actively suppressing background noise from the surrounding environment. This improvement enables the model to extract high-purity human semantic features even in dynamically changing aerial photography backgrounds.

[0043] 3. A distribution-aware coordinate classification and prediction architecture based on DCC is proposed; This invention abandons the traditional direct coordinate regression approach and proposes a Distribution-Aware Coordinate Classification (DCC) architecture. It introduces a Cross-Attention mechanism to process X-axis and Y-axis feature vectors, overcoming the deficiency of existing methods such as SimCC in ignoring the correlation between coordinate axes. The localization task is transformed into a probability distribution classification task on discrete intervals. When a fall causes limb occlusion or blurring, the model can output a "fuzzy correct" result that conforms to the probability distribution, rather than a single "precise incorrect" coordinate, greatly improving the algorithm's robustness.

[0044] 4. Introduce occlusion perception-assisted tasks and KLD soft label supervision; In terms of training strategy, this invention introduces a parallel Aux Head auxiliary branch and a KLD loss function. By using multi-task learning, the model is forced to explicitly learn the "visible / occluded" state of key points, and Gaussian soft labels are used to replace hard labels, enabling the model to tolerate labeling errors during optimization. Weighted optimization is also implemented for challenging samples such as wrists and ankles.

[0045] 5. Establish a multi-dimensional temporal fall detection model with ground-air coordination. This invention no longer relies solely on single-frame image judgment, but combines biomechanical characteristics with a temporal state machine. It comprehensively utilizes torso tilt angle, normalized height drop rate of center of gravity, and aspect ratio features, combined with head movement logic, for false alarm filtering. This logic design effectively distinguishes between "active stretching / tying shoelaces" and "accidental falls," significantly reducing the false alarm rate in practical applications.

[0046] Compared with the prior art, the present invention has the following significant advantages: 1. Significantly improved detection accuracy for "extremely small targets" and from the perspective of UAVs; To address the issue that existing general-purpose models easily lose features of limb extremities at long distances or from the high-altitude perspective of drones, this invention constructs a P2 high-resolution detection branch to physically preserve wrist and ankle features that occupy only a few pixels. Experiments show that this design effectively solves the feature aliasing caused by downsampling, significantly improving the recall rate of the algorithm for key limb points at a 50-meter high-altitude perspective and greatly reducing the missed detections of athletes at a distance.

[0047] 2. Enhanced robustness in complex outdoor environments and occlusion conditions; By introducing a C2PSA spatial attention mechanism into the backbone network and adopting a DCC distributed sensing coordinate classification architecture, this invention effectively overcomes the problems of tree shadow interference and severe limb occlusion during falls on off-road tracks. Compared with traditional regression methods, this invention can still provide smooth predictions that conform to the probability distribution even when occlusion or motion blur occurs, avoiding the phenomenon of coordinates "flying around" and significantly improving positioning accuracy.

[0048] 3. Effectively reduced the false alarm rate and achieved accurate event-level judgment; This invention combines biomechanical characteristics with temporal finite state machines to establish a ground-air coordinated fall detection logic. This logic can not only capture fall actions, but also effectively filter high-frequency interference actions such as tying shoelaces and pre-match stretching through head movement and static dwell detection, ensuring high reliability of alarms.

[0049] 4. It balances high precision and real-time performance, filling a gap in the supervision of outdoor events; All algorithm improvements are based on a lightweight architecture design with low computational load, allowing direct deployment on drone-borne devices or roadside edge computing boxes. The system can achieve millisecond-level real-time inference and alarms, providing all-weather, blind-spot-free safety assurance for wide-area outdoor events such as marathons and trail running, demonstrating extremely high practical value and promising prospects for widespread adoption.

[0050] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A fall detection method for marathon trail running races based on UAVs and attitude estimation, characterized in that, include: Images of marathon trail running events are acquired and processed to obtain the first image; The first image is input into a pre-trained improved YOLOv11 model to obtain the predicted coordinates of the human joints of the runners in the marathon cross-country race. The improved YOLOv11 model introduces a spatial attention mechanism into the YOLOv11 model, adds a detection head and a DCC-based distribution-aware coordinate classification and prediction architecture. Based on the predicted coordinates of the human joints, calculate the trunk tilt angle θ, the normalized height of the center of gravity Hnorm, and the aspect ratio R of the circumscribed rectangle; Based on the torso tilt angle θ, the normalized height of the center of gravity, and the aspect ratio R of the circumscribed rectangle, determine whether a runner in a marathon cross-country race image has fallen.

2. The fall detection method for marathon trail running based on UAV and attitude estimation according to claim 1, characterized in that, Images of marathon trail running events are acquired and processed to obtain the first image, which includes: High-definition network cameras are deployed in accident-prone areas of marathon trail running events to capture images. In blind spots that cannot be covered by fixed equipment, industrial-grade drones equipped with three-axis gimbal cameras are used to capture images. An adaptive filling strategy is used to scale the captured images to obtain the first image.

3. The fall detection method for marathon trail running based on UAV and attitude estimation according to claim 1, characterized in that, The first image is input into the improved YOLOv11 model to obtain the predicted coordinates of the human joints of athletes participating in marathon trail running events, including: The first image is downsampled through the STEM layer to obtain the first feature map. The first feature map is downsampled to obtain the second feature map. The second feature map is downsampled to obtain the third feature map. The third feature map is downsampled to obtain the downsampled third feature map. The downsampled third feature map is processed through a spatial attention mechanism to obtain the fourth feature map. The fourth feature map is downsampled to obtain the downsampled fourth feature map. The downsampled fourth feature map is processed through a spatial attention mechanism to obtain the fifth feature map. The fifth feature map is upsampled to obtain the upsampled fifth feature map. The upsampled fifth feature map is then fused with the fourth feature map to obtain the first fused feature map. The first fused feature map is processed through a DCC-based distribution-aware coordinate classification and prediction architecture to obtain the predicted coordinates of human joints. Specifically, the first fused feature map is processed through a convolutional layer to obtain the X-axis feature vector Vx and the Y-axis feature vector Vy. The correlation matrix between Vx and Vy is calculated, and the correlation matrix is ​​input into the classifier to obtain the probability distribution vector Px on the X-axis and the probability distribution vector Py on the Y-axis. The weighted average of the probability distribution vectors Px on the X-axis and Py on the Y-axis is calculated to obtain the predicted coordinates of human joints.

4. The fall detection method for marathon trail running based on UAV and attitude estimation according to claim 1, characterized in that, The convolutional kernel weights in the improved YOLOv11 model are iteratively updated using the total loss, which is calculated as follows: The first fusion feature map is upsampled to obtain the upsampled first fusion feature map. The upsampled first fusion feature map is then fused with the third feature map to obtain the second fusion feature map. The second fusion feature map is then upsampled to obtain the upsampled second fusion feature map. The upsampled second fusion feature map is then fused with the second feature map to obtain the third fusion feature map. The third fused image is passed through the detection head to obtain the first detection feature map; the first fused feature map is processed through the parallel lightweight auxiliary branch AuxHead to obtain the visibility result, which represents whether the position in the fourth feature map is visible; The visibility loss is calculated based on the joint optimized location loss function and the visibility results. Based on the true coordinates of the key points and the preset variance σ, a Gaussian distribution is generated as a soft label for the predicted coordinates. The KL divergence between the model's predicted coordinates and the soft label is calculated to obtain the KL divergence loss. The visibility loss and the KL divergence loss together constitute the total loss.

5. The fall detection method for marathon trail running based on UAV and attitude estimation according to claim 1, characterized in that, Based on the predicted coordinates of the human joints, calculate the trunk tilt angle θ, the normalized height of the center of gravity Hnorm, and the aspect ratio R of the circumscribed rectangle, including: In the predicted coordinates of human joints, connect the key points of the neck and the mid-hip point, calculate the angle between the connected vector and the vertical line of gravity to obtain the torso tilt angle θ; obtain the Y coordinate of the mid-hip point, calculate the relative height of the Y coordinate of the mid-hip point from the ground, divide the relative height by the height of the human detection box to obtain the normalized height of the center of gravity Hnorm; calculate the ratio of the width W to the height H of the human detection box to obtain the aspect ratio R of the bounding rectangle.

6. The fall detection method for marathon trail running based on UAV and attitude estimation according to claim 1, characterized in that, Based on the torso tilt angle θ, the normalized height of the center of gravity, and the aspect ratio R of the circumscribed rectangle, determine whether a runner in a marathon trail running image has fallen, including: When θ > 45° and the normalized height of the center of gravity Hnorm decreases by more than a preset threshold ΔH within the time window, the runners in the image representing the marathon trail running event are in an unstable state. When an athlete in a marathon trail running image is in an unstable state and meets any of the target conditions, it indicates that the athlete in the marathon trail running image is in a fall state. When an athlete in a marathon cross-country race image is in a fallen state, if the head key point in the predicted coordinates of the human joints exhibits vertical reciprocating motion with an amplitude exceeding a preset threshold ΔH within a preset time period, the athlete's state in the marathon cross-country race image is reset to a normal state. If the average optical flow velocity of all key points in the predicted coordinates of the human joints is lower than the static threshold, and the fallen state is maintained for more than the safety confirmation time, the athlete in the marathon cross-country race image is confirmed to have fallen.

7. The fall detection method for marathon trail running based on UAV and attitude estimation according to claim 1, characterized in that, The target conditions include: the aspect ratio R of the circumscribed rectangle is greater than 1.2; the torso tilt angle θ is greater than 65° and the duration exceeds the preset duration.

Images