A method and system for assessing fall risk in the elderly based on depth cameras

Abstract

This invention discloses a method and system for assessing fall risk in the elderly based on a depth camera, specifically relating to the field of medical image data processing and risk assessment. The method includes acquiring a depth image sequence corresponding to the subject's assessment actions, identifying the three-dimensional coordinates of human bone points and the toes, heels, ankles, and knees of both feet frame by frame to form a bone point sequence, dividing the human body into segments based on the bone point sequence, and calculating the coordinates of the human centroid based on the segment mass value and centroid position of each segment. By extracting bone points from the depth images, reconstructing the dynamic support boundary of both feet, and calculating the risk distance of the human centroid projection relative to the dynamic support boundary, this invention addresses the problems of insufficient information on the outer edges of the feet under single-depth-camera acquisition conditions and continuous changes in support status during movement, making it difficult for the support boundary to correspond to the actual foot bearing capacity and for the risk distance of the human centroid projection relative to the support boundary to be stably obtained.

Description

Technical Field

[0001] This invention relates to the field of medical image data processing and risk assessment technology, and more specifically, to a method and system for assessing fall risk in the elderly based on a depth camera. Background Technology

[0002] In fall risk assessment for the elderly, the main approach of existing technologies is to replace questionnaire judgment and manual observation with quantifiable biomechanical indicators. In engineering, the human body's center of mass and foot support information are usually obtained by using a three-dimensional motion capture device and a force measuring platform, or by extracting human skeletal landmarks through a depth camera and then calculating the support range of both feet by combining the positions of the toes and heels. The risk assessment result is then given based on the positional relationship between the two-dimensional projection of the human body's center of mass and the support boundary. In hospital outpatient screening, rehabilitation reassessment, and daily assessment scenarios in elderly care institutions, the system typically needs to simultaneously meet the requirements of simple equipment deployment, short assessment time, no additional devices worn by subjects, and doctors being able to directly view the assessment process and results in a virtual reality interface. Under these conditions, although depth cameras can continuously acquire human posture information, their ability to acquire the outer contours of the feet is insufficient. Furthermore, the orientation and actual bearing range of the feet of elderly people continuously change during starting, stopping, turning, and sitting-standing transitions. This leads to repeated problems such as the display results of the support boundary not corresponding to the actual foot support range, abnormal fluctuations in the distance value of the human centroid projection relative to the support boundary at adjacent moments, and inconsistencies between some risk conclusions and the actual movement state of the same subject. Since the existing solutions still rely mainly on static estimation to supplement the missing information of the outer edge of the feet, and have not formed a support boundary reconstruction process that combines continuous movement state and pattern recognition, the support range on which the risk distance calculation is based becomes distorted during movement switching. Therefore, under single-depth camera acquisition conditions, given the lack of information on the outer edges of both feet and the continuous changes in support state during movement, how to reconstruct the dynamic support boundary corresponding to the actual foot bearing range, and thereby stably obtain the risk distance of the human body's center of mass projection relative to the support boundary, becomes the technical problem to be solved in this application. Summary of the Invention

[0003] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a method and system for assessing fall risk in the elderly based on a depth camera. This method solves the problems mentioned in the background art by extracting bone points from depth images, reconstructing the dynamic support boundary of both feet, and calculating the risk distance of the human centroid projection relative to the dynamic support boundary.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for assessing fall risk in the elderly based on a depth camera, comprising: S1. Obtain the depth image sequence corresponding to the subject's assessment action, and identify the three-dimensional coordinates of human bone points and the toe points, heel points, ankle points and knee points of both feet frame by frame to form a bone point sequence. S2. Divide the human body into segments according to the bone point sequence, calculate the human body centroid coordinates according to the segment mass value and centroid position of each segment, and project the human body centroid coordinates onto the ground plane to form a centroid projection sequence. S3. Construct the foot axis based on the toe and heel points of each foot in the bone point sequence, construct the lower leg axis based on the corresponding knee and ankle points, divide the foot axis into six equal parts to obtain five points, determine the lateral direction of the foot by the cross product direction of the lower leg axis and the foot axis, and determine five outer edge points of the foot based on the length correction amount corresponding to each of the five points to form the outer edge sequence of the foot. S4. Calculate the contact coefficients of the two feet based on the position changes of the two feet between adjacent frames in the bone point sequence. Scaling the positions of the toe points, heel points and outer edge points of the two feet according to the contact coefficients to form an effective boundary point set. Perform convex hull calculation on the effective boundary point set to form a dynamic support boundary sequence. S5. Based on the centroid projection sequence and the dynamic support boundary sequence, calculate the directed distance from the centroid projection point to each segment of the dynamic support boundary frame by frame. Sort the directed distances corresponding to each segment in ascending order of absolute value and take the first directed distance as the boundary distance of the current frame. Then, accumulate the boundary distances of each frame with positive boundary distances in chronological order to form the fall risk value.

[0005] In a preferred embodiment, S1 includes: S1-1. Convert each frame of depth image in the depth image sequence into a three-dimensional coordinate point set, extract the human body connected region and human body contour, arrange the height coordinates of each point of the human body contour in ascending order, take the contour point corresponding to the first position of the sort to form the foot point set, and then separate the foot point set into the left foot point set and the right foot point set according to the human body midline to form the foot point set sequence. S1-2. For each foot in the foot point set sequence, calculate the distance between each pair of points in the foot point set, arrange the distances between the points in ascending order, take the pair of points corresponding to the last position in the sort to form the foot axis, determine the end closer to the human body's central axis as the heel point, determine the end farther from the human body's central axis as the toe point, and then determine the intersection of the midpoint of the foot axis along the height direction with the human body's central axis as the ankle point, thus forming the foot point sequence. S1-3. Extract the center line from the connected regions of the human body in each frame, determine the lower limb inflection point above the ankle point in the center line as the knee point, and write the bifurcation point and endpoint of the center line into the human bone points to form a bone point sequence.

[0006] In a preferred embodiment, S2 includes: S2-1. Divide the body into trunk, thigh, calf and foot segments according to the connection relationship between adjacent bone points in the bone point sequence. Select multiple sampling points at equal intervals along the segment axis corresponding to each human body segment to form a sampling point sequence. S2-2. For each sampling point in the sampling point sequence, find the intersection with the human body contour along the vertical direction of the segment axis to obtain the left and right intersection points. Calculate the distance between the left and right intersection points, and multiply the distance corresponding to each sampling point with the position distance of the sampling point on the segment axis. Then, accumulate the results according to the human body segments to form a segmented quality value sequence.

[0007] In a preferred embodiment, S2 further includes: S2-3. For each sampling point in the sampling point sequence, multiply the position distance of each sampling point on the segment axis by the corresponding distance and accumulate them according to the human body segment. Then divide the accumulated result by the accumulated distance value corresponding to the human body segment to form the segment centroid position sequence. S2-4. Determine the centroid coordinates of each human body segment based on the segmented mass value sequence and the segmented centroid position sequence. Then, weight and accumulate the centroid coordinates of each human body segment according to the segmented mass value sequence to obtain the centroid coordinates of the human body. Finally, project the centroid coordinates of the human body onto the ground plane to form a centroid projection sequence.

[0008] In a preferred embodiment, S3 includes: S3-1. Construct the foot axis based on the toe and heel points of each foot in each frame of the bone point sequence, and construct the lower leg axis based on the corresponding knee and ankle points. Divide the foot axis into six equal parts to obtain five sub-points, and calculate the projection value of the cross product of the lower leg axis and the foot axis onto the normal direction of the ground plane to form a direction value sequence. S3-2. For the current frame in the direction value sequence, extract the direction value of the previous frame, the direction value of the current frame, and the direction value of the next frame. Compare the signs of the direction values ​​of the previous frame and the current frame, and the direction values ​​of the current frame and the next frame. Calculate the absolute differences between the direction values ​​of the current frame and the previous frame, and the absolute differences between the direction values ​​of the current frame and the next frame. If the two sign comparison results are consistent and the ranking positions corresponding to the differences of the two absolute values ​​are both at the first position, the side corresponding to the current frame direction value is determined as the outer side. If the two sign comparison results are inconsistent and the signs of the direction values ​​of the previous frame and the next frame are consistent, the two sides of the foot axis are determined as the first candidate side and the second candidate side, respectively. If the two sign comparison results are consistent and the ranking positions corresponding to the differences of the two absolute values ​​are not both at the first position, the side corresponding to the direction value of the previous frame and the side corresponding to the direction value of the next frame are determined as the side to be detected.

[0009] In a preferred embodiment, S3 further includes: S3-3. For the first candidate side, the second candidate side, or the side to be tested, retrieve the first confirmed frame in chronological order where the outer edge of the foot has been formed, and retrieve the first confirmed frame in chronological order where the outer edge of the foot has been formed. On each side, determine five base points at five points according to 50 mm for males and 35 mm for females. Calculate the cumulative distance between the five base points corresponding to each side and the corresponding outer edge of the foot in the previous confirmed frame, the cumulative distance between the five base points corresponding to each side and the corresponding outer edge of the foot in the subsequent confirmed frame, and the difference between the cumulative distance between the previous and subsequent confirmed frames. Then, sort the sum of the cumulative distance between the previous and subsequent confirmed frames and the difference between the cumulative distance between the previous and subsequent confirmed frames in ascending order. Take the side that is first in both sorts and determine it as the outer side.

[0010] In a preferred embodiment, S3 further includes: S3-4. Determine the lateral direction of the foot based on the lateral side, calculate the angle between the lower leg axis and the foot axis in the current frame, the angle between the lower leg axis and the foot axis in the previous confirmation frame, and the angle between the lower leg axis and the foot axis in the subsequent confirmation frame. Distribute the difference between the average of the angle in the current frame, the angle in the previous confirmation frame, and the angle in the subsequent confirmation frame to five points according to the ratio of the distance from the heel point to the length of the foot axis, forming a length correction sequence. S3-5. Along the outer side of the foot, locate the five points as five outer edge points of the foot according to the sum of the length correction amount corresponding to the basic length and the length correction amount sequence, and collect the five outer edge points of the foot in time order to form the outer edge sequence of the foot.

[0011] In a preferred embodiment, S4 includes: S4-1. For the toe point, heel point and outer edge point of the feet in two adjacent frames in the bone point sequence, calculate the distance between each point in the ground plane, and accumulate the distance of all points corresponding to each foot to form the foot displacement value. Then, sort the foot displacement values ​​corresponding to both feet in ascending order, take the first foot displacement value in the sorted order and divide it one by one with the foot displacement values ​​corresponding to each foot to form the contact coefficient corresponding to both feet. S4-2. Using the ankle points of each foot as the base points, keep the direction of the lines connecting the toe points, heel points, and outer edge points of each foot to the ankle points unchanged, and scale the length of each line according to the contact coefficient to form an effective set of boundary points. S4-3. Project the effective boundary point set onto the ground plane, arrange the projected points in ascending order of their horizontal coordinates and perform convex hull calculation. The convex hull calculation is as follows: calculate the turning values ​​of three adjacent points in sequence, delete the middle point located on the inner side, retain the outer points and connect them end to end to form a dynamic support boundary sequence.

[0012] In a preferred embodiment, S5 includes: S5-1. For each centroid projection point in the centroid projection sequence, connect adjacent boundary points in the dynamic support boundary sequence to form each edge segment. Calculate the position of the perpendicular foot of the centroid projection point on the straight line of each edge segment. When the perpendicular foot is inside the edge segment, take the distance from the centroid projection point to the perpendicular foot as the directed distance of the edge segment. When the perpendicular foot is outside the edge segment, take the distance from the centroid projection point to the shortest endpoint among the edge segment endpoints as the directed distance of the edge segment. Record the directed distance corresponding to the centroid projection point being inside the dynamic support boundary as a negative value and the directed distance corresponding to the centroid projection point being outside the dynamic support boundary as a positive value. S5-2. Sort the directed distances corresponding to each edge segment in ascending order of absolute value, take the first directed distance in the sorted order as the current frame boundary distance, and collect the current frame boundary distances in frame order to form a boundary distance sequence. S5-3. The boundary distances of each frame with a positive boundary distance in the boundary distance sequence are accumulated in chronological order to form a fall risk value.

[0013] In a preferred embodiment, a depth camera-based fall risk assessment system for the elderly includes: The image point acquisition module is used to acquire depth image sequences corresponding to the subject's assessment actions, and to identify the three-dimensional coordinates of human bone points and the toe points, heel points, ankle points and knee points of both feet frame by frame to form a bone point sequence; The centroid determination module divides the human body into segments based on the bone point sequence, calculates the centroid coordinates of the human body according to the segment mass value and centroid position of each segment, and projects the centroid coordinates of the human body onto the ground plane to form a centroid projection sequence. The outer edge positioning module constructs the foot axis based on the toe and heel points of each foot in the bone point sequence, constructs the lower leg axis based on the corresponding knee and ankle points, divides the foot axis into six equal parts to obtain five sub-points, determines the lateral direction of the foot by the cross product direction of the lower leg axis and the foot axis, and determines five outer edge points of the foot based on the length correction amount corresponding to each of the five sub-points, thus forming the outer edge sequence of the foot. The boundary generation module calculates the contact coefficients of the two feet based on the position changes of the two feet between adjacent frames in the bone point sequence, scales the positions of the toe points, heel points and outer edge points of the two feet according to the contact coefficients to form an effective boundary point set, and performs convex hull calculation on the effective boundary point set to form a dynamic support boundary sequence. The risk calculation module calculates the directed distance from the centroid projection point to each segment of the dynamic support boundary frame by frame based on the centroid projection sequence and the dynamic support boundary sequence. The directed distances corresponding to each segment are sorted in ascending order of absolute value, and the directed distance at the top of the sort is taken as the boundary distance of the current frame. The boundary distances of each frame with a positive boundary distance are then accumulated in time order to form the fall risk value.

[0014] The technical effects and advantages of this invention are as follows: By extracting bone points from depth images and calculating the human centroid, and then combining foot lateral edge completion, contact coefficient scaling and dynamic support boundary reconstruction, the distance calculation of the centroid projection relative to the support boundary can be established on a boundary basis that is more consistent with the actual foot bearing range under single depth camera conditions. This can relatively improve the problem of abnormal fluctuations in risk distance and inconsistent evaluation conclusions during action switching. By dividing the foot axis into six equal parts, determining the lateral side based on the direction relationship of the lower leg axis, and locating five foot outer edge points point by point according to the length correction amount, a foot outer edge sequence corresponding to the current foot posture can be formed under the condition of missing foot outer edge contour, which can relatively improve the distortion of support range caused by static extrapolation. By extracting sampling points according to human body segments, calculating segment mass values ​​and segment centroid positions, and then summarizing the human body centroid coordinates and projecting them onto the ground plane, the centroid position calculation can establish a correspondence with the human body contour distribution in the current frame, which can relatively improve the adaptation deficiencies caused by directly applying a fixed human body proportion. By forming a contact coefficient based on the positional changes of the two feet between adjacent frames, and scaling the positions of the toe point, heel point, and outer edge point of the foot according to the contact coefficient, and then performing convex hull calculation to form a dynamic support boundary sequence, the support boundary can be made to change synchronously with the degree of load participation, thereby suppressing the expansion effect of the swinging foot on the support range. By calculating the directed distance from the centroid projection point to each segment of the dynamic support boundary frame by frame, extracting the boundary distance of the current frame, and accumulating the positive boundary distances in time sequence to form a fall risk value, the degree of balance boundary violation can be transformed into a continuous quantitative result, which makes it easier to intuitively display and compare the risk changes of different assessment processes in the virtual reality interface. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method steps of the present invention.

[0016] Figure 2 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Refer to the instruction manual appendix Figure 1-2The present invention provides a method for assessing fall risk in the elderly based on a depth camera, comprising: S1. Obtain the depth image sequence corresponding to the subject's assessment action, and identify the three-dimensional coordinates of human bone points and the toe points, heel points, ankle points and knee points of both feet frame by frame to form a bone point sequence. In this embodiment, S1 is used to extract the basic bone point data required for subsequent calculations from the depth image sequence. The processing order is as follows: first, each frame of depth image is restored to a three-dimensional coordinate point set; then, the human body connected region and human body contour are extracted from the three-dimensional coordinate point set, and the foot point set corresponding to both feet is determined at the bottom of the human body contour; then, the heel point, toe point and ankle point are extracted for the foot point set of each foot; finally, the knee point and other human body bone points are extracted from the human body connected region, thereby forming a bone point sequence shared by subsequent human body segmentation, foot outer edge reconstruction and dynamic support boundary calculation. The implementation process includes the following steps: First, the depth images of each frame in the depth image sequence are converted into a set of three-dimensional coordinate points according to the intrinsic parameter matrix of the depth camera and the correspondence between pixel coordinates. Then, connected regions are constructed according to the spatial adjacency relationship between the points, and the connected region with the first point in the sorted order is taken as the human body connected region. After the human body connected region is determined, the human body contour is extracted along its outer edge, the height coordinates of each point of the human body contour are read and sorted in ascending order, and all contour points with the same height coordinates as the first point in the sorted order are taken to form the foot point set. Then, the line connecting the center projection of the human body connected region on the ground plane is taken as the human body central axis, and the foot point set is separated into the left foot point set and the right foot point set according to the left and right sides of the human body central axis, thus forming a foot point set sequence frame by frame. Furthermore, for each foot in the foot point set sequence, the distance between each pair of points in the foot point set is calculated, and all the distances are arranged in ascending order. The pair of points corresponding to the last point in the sort is taken as the two endpoints of the foot, and the foot axis is formed by these two endpoints. Then, the distance from these two endpoints to the human body midline is calculated. The end with the shorter distance is determined as the heel point, and the end with the longer distance is determined as the toe point. After the heel point and toe point are determined, the midpoint of the foot axis is taken as the foot midpoint, and a straight line is drawn along the height direction. The intersection of this straight line and the center line of the lower limb region in the current frame is taken as the ankle point. Thus, the foot point sequence is formed frame by frame. When there is a front-back misalignment between the left and right feet, the above-mentioned distance calculation and endpoint determination are still performed independently in the left foot point set and the right foot point set, respectively, so that the heel point, toe point, and ankle point always correspond to the position of the corresponding foot. After obtaining the foot point sequence, when extracting the center line of the human body connected region in each frame, the human body connected region is first refined into a skeleton line with a single pixel width, and then the bifurcation point, endpoint, and inflection point are extracted along the skeleton line. Among them, the inflection point is formed by the corner change of adjacent skeleton line segments, and the inflection point located above the ankle point and on the branch of the lower limb skeleton is taken as the knee point. After the knee point is determined, the bifurcation point and endpoint of the center line are written into the human bone point, and the toe point, heel point, ankle point in the foot point sequence and the knee point of the current frame are written into the human bone point set of the current frame. Finally, the human bone point sets of each frame are collected in chronological order to form a bone point sequence. When there are multiple inflection points on the same lower limb, the inflection point that is first in the distance from the ankle point along the center line and located below the hip is taken as the knee point to ensure that the determination rules of the knee point are consistent between adjacent frames. Through the above processing, the depth image sequence can be stably converted into a bone point sequence that can be directly called in subsequent steps, and the correspondence between the foot point set, foot point sequence and bone point sequence can be kept consistent, thereby eliminating the unclear problems in foot extraction, ankle point localization and knee point recognition, which facilitates the subsequent calculation of human body centroid, determination of the lateral direction of the foot and generation of dynamic support boundary based on bone point sequence. In practical applications: When the subject completes standing, starting or turning movements in the assessment area, the depth camera continuously outputs multiple frames of depth images. Each frame of depth image is first converted into a three-dimensional coordinate point set, then the human body connected region and human body contour are extracted, and the left foot point set and right foot point set are obtained from the bottom of the human body contour. Then, the heel point, toe point and ankle point are extracted from the foot point set corresponding to each foot, and the knee point and other human body bone points are extracted from the center line of the human body connected region. Finally, a bone point sequence arranged in chronological order is formed for direct use in subsequent steps.

[0019] S2. Divide the human body into segments according to the bone point sequence, calculate the human body centroid coordinates according to the segment mass value and centroid position of each segment, and project the human body centroid coordinates onto the ground plane to form a centroid projection sequence. In this embodiment, S2 is used to further obtain the human centroid coordinates and centroid projection sequence that can be used for fall risk assessment based on the bone point sequence that has been determined. The processing idea is not to directly apply a fixed human body proportion, but to calculate the segment mass value and segment centroid position of each human body segment according to the distribution of the human body contour in each human body segment in the current frame. Then, the segment centroid coordinates of each human body segment are summed together to obtain the human centroid coordinates. Finally, the human centroid coordinates are projected onto the ground plane to form the centroid projection sequence. The trunk, thigh, lower leg, and foot segments are all formed by connecting adjacent bone points in a bone point sequence. Sampling points, left and right intersections, segment mass values, and segment centroid positions are obtained step by step based on this connection relationship. The implementation process includes the following steps: First, the body is divided into trunk, thigh, calf, and foot segments based on the connection relationships between adjacent bone points in the bone point sequence. The trunk segment is formed by connecting the trunk point and the hip point; the thigh segment by connecting the hip point and the knee point; the calf segment by connecting the knee point and the ankle point; and the foot segment by connecting the ankle point and the toe point. After each body segment is determined, the line connecting the two ends of each segment's bone points is used as the segment axis. Multiple sampling points are then selected along each segment axis at equal intervals. The equal interval method involves dividing the segment axis length into several segments of equal length and taking each segment point as a sampling point, thus forming a sampling point sequence arranged according to the body segments. In this embodiment, the same number of sampling points are used for each body segment, providing a clear basis for calculating the distance between the dynamic support boundary and the centroid projection. Furthermore, for each sampling point in the sampling point sequence, a cross-section perpendicular to the segment axis is constructed at the location of the sampling point. This cross-section intersects with the human body contour of the current frame to obtain the left and right intersection points on both sides of the segment axis. The distance between the left and right intersection points is then calculated as the human body lateral dimension corresponding to the sampling point. Subsequently, the path length from the sampling point to the starting point of the corresponding human body segment along the segment axis is calculated. This path length is multiplied by the corresponding human body lateral dimension, and then accumulated item by item for all sampling points within the same human body segment to form the segment mass value of the human body segment. Based on this, the same calculation is performed sequentially on all human body segments to form a segment mass value sequence corresponding one-to-one with the torso segment, thigh segment, calf segment, and foot segment. The segment mass value is used to characterize the relative mass distribution of each human body segment in the current frame and does not represent the actual weighing result. After obtaining the segmented mass value sequence, for each sampling point in the sampling point sequence, the path length from the sampling point to the corresponding human segment starting point along the segment axis and the corresponding human lateral dimension of the sampling point are read. The two are multiplied and accumulated by human segment to obtain the weighted position cumulative value of the human segment. Then, the human lateral dimensions corresponding to all sampling points in the human segment are accumulated item by item to obtain the size cumulative value of the human segment. The weighted position cumulative value is divided by the size cumulative value to obtain the segment centroid position of the human segment along the segment axis. After that, the above calculation is performed on all human segments in sequence to form a segment centroid position sequence. In this embodiment, the segment centroid position represents the distance from the segment centroid point to the human segment starting point along the segment axis, so it can be directly used for the subsequent segment centroid coordinate positioning. After determining the segmented mass value sequence and segmented centroid position sequence for each human body segment, the starting point of each human body segment is used as the base point. The segmented centroid point is located along the corresponding segmented axis according to the segmented centroid position of the human body segment, and the segmented centroid coordinates of each human body segment are determined accordingly. Then, the segmented centroid coordinates of each human body segment are multiplied by the corresponding segmented mass value, and the product results are accumulated item by item. The sum of the accumulated results is divided by the sum of all segmented mass values ​​to obtain the human body centroid coordinates. After the human body centroid coordinates are determined, the human body centroid coordinates are projected onto the ground plane along the normal direction of the ground plane to obtain the centroid projection point of the current frame. Then, the centroid projection points of each frame are collected in chronological order to form a centroid projection sequence. The ground plane is the plane where the foot point set is located, or the plane obtained by fitting the ground region in the depth image. It remains unchanged in the same evaluation process to ensure that the centroid projection results of each frame are under the same reference plane. Through the above processing, under the condition that the bone point sequence and human body contour are known, the segment mass value and segment centroid position of each human body segment can be obtained segment by segment, and the human body centroid coordinates and centroid projection sequence corresponding to the current action state can be formed accordingly. This avoids the problem of insufficient adaptation caused by directly using a fixed ratio, and also provides a clear data basis for the subsequent calculation of the distance between the dynamic support boundary and the centroid projection. In practical applications: After the subject completes the standing, stepping, or turning movements, the body can be divided into trunk, thigh, calf, and foot segments according to the bone point sequence. Sampling points are then selected at equal intervals on each body segment, and the corresponding lateral dimensions of the human body are obtained from the intersection of the sampling points and the human body contour. Subsequently, the segmental mass value and segmental centroid position of each body segment are calculated. Then, the segmental centroid coordinates of each body segment are located and summarized to obtain the human body centroid coordinates. Finally, the human body centroid coordinates are projected onto the ground plane to form a centroid projection sequence that runs through the entire assessment process, which can be directly called upon for subsequent fall risk value calculations.

[0020] S3. Construct the foot axis based on the toe and heel points of each foot in the bone point sequence, construct the lower leg axis based on the corresponding knee and ankle points, divide the foot axis into six equal parts to obtain five points, determine the lateral direction of the foot by the cross product direction of the lower leg axis and the foot axis, and determine five outer edge points of the foot based on the length correction amount corresponding to each of the five points to form the outer edge sequence of the foot. In this embodiment, S3 is used to further obtain the outer edge point of each foot in each frame, based on the already determined toe point, heel point, knee point, and ankle point. The processing is divided into two levels: first, the lateral side is determined based on the spatial relationship between the foot axis and the lower leg axis; then, based on the lateral side, the distribution of the outer edge points of the foot in the previous and subsequent confirmation frames, and the angle change in the current frame, the extension of the five sub-points in the lateral direction of the foot is corrected point by point, thereby forming the outer edge sequence of the foot. Among them, the direction value sequence is used to characterize the relative lateral relationship between the foot orientation and the lower leg orientation in the current frame; the lateral side is used to determine which side of the foot axis corresponds to the lateral direction of the foot; the length correction sequence is used to determine the extension correction amount of the five sub-points beyond the basic length; and the previous and subsequent confirmation frames are used to provide front and back references when the orientation determination in the current frame is unstable. The implementation process includes the following steps: First, a foot axis is constructed based on the toe and heel points of each foot in each frame of the bone point sequence, with the heel point as the starting point and the toe point as the ending point. Then, a lower leg axis is constructed based on the corresponding knee and ankle points, with the ankle point as the starting point and the knee point as the ending point. After the two axes are determined, the foot axis is divided into six segments of equal length, and five points (excluding the heel and toe points) from each segment are taken as five sub-points. Next, the normal to the ground plane is used as the direction reference, and the projection value of the cross product of the lower leg axis and the foot axis onto the normal to the ground plane is calculated. This projection value is then used as the direction value for the corresponding foot in the current frame. The foot axis direction is fixed from the heel point to the toe point, and the ground plane normal remains unchanged throughout the same evaluation process; therefore, the positive or negative direction of the direction value stably corresponds to the two sides of the foot axis. Finally, the direction values ​​corresponding to each foot in each frame are collected in chronological order to form a direction value sequence. After the direction value sequence is formed, for any current frame, extract the direction values ​​of the preceding frame, the current frame, and the following frame that are adjacent to it in time. Compare the signs of the direction values ​​of the preceding and current frames, and the signs of the direction values ​​of the current and following frames. Simultaneously, calculate the absolute differences between the current and preceding frame direction values, and the absolute differences between the current and following frame direction values, and sort these two absolute differences in ascending order. If the two sign comparisons are consistent, and the sorted positions corresponding to the two absolute differences are both at the first position, it indicates that the current frame direction value is on the same side as the preceding and following frames and the change is small. In this case, the side corresponding to the current frame direction value is directly determined as the outer side. If the two sign comparisons are inconsistent... If the signs of the direction values ​​in the previous and subsequent frames are consistent, it indicates that the current frame is in the side-switching interval or local disturbance interval. In this case, the outer side is not directly determined, but the two sides of the foot axis are recorded as the first candidate side and the second candidate side, respectively. If the signs of the two comparisons are consistent, but the ranking positions corresponding to the difference in the absolute values ​​of the two items are not simultaneously at the first position, it indicates that although the current frame is on the same side as the previous and subsequent frames, its direction value changes more significantly. In this case, the side corresponding to the direction value of the previous frame and the side corresponding to the direction value of the subsequent frame are recorded as the side to be detected. When the previous frame cannot be obtained in the first frame of the sequence, the subsequent frame and the current frame are used to form a comparison group. When the subsequent frame cannot be obtained in the last frame of the sequence, the previous frame and the current frame are used to form a comparison group to ensure that the same judgment process can still be performed on the boundary frames. After the first candidate side, second candidate side, or the side to be tested is determined, the process continues to search forward in time for the first confirmed frame with a formed outer edge point, and backward in time for the first confirmed frame with a formed outer edge point. If the current frame is at the beginning of the sequence and only a confirmed frame is found, then the confirmed frame is used as the sole reference frame. If the current frame is at the end of the sequence and only a confirmed frame is found, then the confirmed frame is used as the sole reference frame. After the reference frame is determined, for each candidate side, five base points are determined along the direction corresponding to the candidate side, starting from five points, at a ratio of 50 mm for males or 35 mm for females. Then, the point distances between the five base points corresponding to the candidate side and the five outer edge points corresponding to the previous confirmed frame are calculated one by one, and the five point distances are accumulated to form the accumulated point distance value of the previous confirmed frame. Then, the distances between the candidate side and the five outer edge points of the previous confirmed frame are calculated one by one. The distances between the five base points corresponding to the lateral side and the five foot outer edge points corresponding to the subsequent confirmed frame are calculated, and the five distances are accumulated to form the accumulated distance value of the subsequent confirmed frame. Then, the difference between the accumulated distance value of the previous confirmed frame and the accumulated distance value of the subsequent confirmed frame is calculated, and the sum of the accumulated distance values ​​of the previous confirmed frame and the subsequent confirmed frame is also calculated. Finally, the sums of the accumulated distance values ​​corresponding to each candidate lateral side are sorted in ascending order, and the differences of the accumulated distance values ​​corresponding to each candidate lateral side are sorted in ascending order. The candidate lateral side that ranks first in both sorts is determined as the outer lateral side. When there is only one reference frame, the candidate lateral side that ranks first in the accumulated distance value corresponding to that reference frame is directly determined as the outer lateral side. Through this process, the outer lateral side of the current frame can still be corrected using the foot outer edge distribution of the previous and subsequent confirmed frames when the direction value of the current frame is unstable. After determining the lateral side, the lateral direction of the foot is determined based on the lateral side. Specifically, if the lateral side corresponds to the left side of the foot axis, the direction from the dividing point to the left side of the foot axis is recorded as the lateral direction of the foot; if the lateral side corresponds to the right side of the foot axis, the direction from the dividing point to the right side of the foot axis is recorded as the lateral direction of the foot. After determining the lateral direction of the foot, the angle between the lower leg axis and the foot axis in the current frame, the angle between the lower leg axis and the foot axis in the previous confirmation frame, and the angle between the lower leg axis and the foot axis in the subsequent confirmation frame are calculated respectively. The angle difference is obtained by subtracting the average of the angles in the previous and subsequent confirmation frames from the angle in the current frame. When only one exists... When there are five reference frames, the angle difference is obtained by subtracting the angle of the reference frame from the angle of the current frame. Then, the ratio of the distance from the five points to the heel point to the length of the foot axis is calculated, and the five ratios are used as the allocation coefficients for the five points. The angle difference is then allocated to the five points according to the five allocation coefficients to form a length correction sequence. When the angle difference is positive, the length correction is written outward along the outer side of the foot. When the angle difference is negative, the length correction is written inward along the outer side of the foot. This allows the length correction sequence to reflect the local expansion changes of the outer edge of the foot in the current frame relative to the previous and previous reference frames. After obtaining the length correction sequence, five sub-points are located along the outer side of the foot according to the sum of the base length and the corresponding length correction in the length correction sequence, resulting in five outer edge points of the foot. The base length is consistent with the length used when determining the base point: 50 mm for males and 35 mm for females. The final extension of each sub-point is obtained by adding the base length and the corresponding length correction. When the length correction corresponding to a sub-point is negative, and the sum of the base length and the length correction is still greater than zero, the outer edge point of the foot is located according to the length corresponding to that sum. When the sum of the base length and the length correction is equal to zero, the outer edge point of the foot coincides with that sub-point. After the five outer edge points of the foot are determined, the five outer edge points of each foot in each frame are collected in chronological order to form the outer edge sequence of the foot. In this embodiment, each foot in each frame of the outer edge sequence corresponds to five outer edge points, and the five outer edge points correspond one-to-one with the five sub-points, so that they can be directly used for the construction of the effective boundary point set and the dynamic support boundary. Through the above processing, even when the depth camera cannot directly obtain the complete outer contour of the foot, five outer edge points of the foot can be formed frame by frame based on the foot axis, lower leg axis, front and back reference frames, and angle difference allocation results. This makes the determination process of the outer edge of the foot have clear lateral basis, front and back connection relationship, and point-by-point correction rules. It avoids directly calling the outer direction of the foot as an existing quantity, and also avoids the problem of inconsistency with the actual foot bearing range caused by the five outer edge points being determined by only a fixed outer extension. In the subsequent dynamic support boundary calculation, the boundary basis corresponding to the current frame action state can be maintained. In practical applications: When a subject completes a starting, ending, or turning motion, the foot axis, lower leg axis, and direction value sequence are first obtained based on the bone point sequence of each frame. Then, the lateral side of the current frame is determined by comparing the direction values ​​of the current frame, the previous frame, and the next frame. When the direction value of the current frame is stable, the lateral side is directly determined. When the direction value of the current frame is in the switching interval, the previous and next confirmation frames are retrieved, and the lateral side of the current frame is selected by accumulating the distances between the five base points corresponding to each candidate side and the foot outer edge points of the previous and next confirmation frames. Afterward, a length correction amount is allocated based on the angle difference between the lower leg axis of the current frame and the previous and next confirmation frames, and the five foot outer edge points are located point by point along the lateral direction of the foot, finally forming a foot outer edge sequence arranged in chronological order, which is used for the generation of the subsequent dynamic support boundary sequence.

[0021] S4. Calculate the contact coefficients of the two feet based on the position changes of the two feet between adjacent frames in the bone point sequence. Scaling the positions of the toe points, heel points and outer edge points of the two feet according to the contact coefficients to form an effective boundary point set. Perform convex hull calculation on the effective boundary point set to form a dynamic support boundary sequence. In this embodiment, S4 is used to determine the degree of foot bearing participation in the current frame based on the positional changes of the feet between adjacent frames, and to form a dynamic support boundary sequence corresponding to the foot bearing range of the current frame. The processing procedure is as follows: First, the foot displacement values ​​of the two feet are calculated based on the positional changes of the toe points, heel points, and outer edge points of the feet in two adjacent frames, and then the contact coefficients of the two feet are formed from the foot displacement values. Subsequently, the length of the toe points, heel points, and outer edge points of the two feet is scaled using the ankle points of each foot as the base points to form an effective boundary point set. Finally, the effective boundary point set is projected onto the ground plane, and convex hull calculation is performed in the ground plane to obtain the dynamic support boundary of the current frame. The implementation process includes the following steps: First, for the toe points, heel points, and outer edge points of both feet in adjacent frames of the bone point sequence, the corresponding points in the previous and current frames are projected onto the ground plane, and the Euclidean distance between the corresponding two points is calculated. This distance is used as the point distance between adjacent frames. The outer edge points are the five points in the outer edge sequence that correspond one-to-one with the five sub-points. Therefore, for each foot, a toe point distance, a heel point distance, and the distances of the five outer edge points are calculated between adjacent frames. After determining the point distances, all points corresponding to the same foot are... The distances are accumulated item by item to form the foot displacement value of the foot in the current frame. Then, the foot displacement values ​​corresponding to the left and right feet are sorted in ascending order, and the foot displacement value corresponding to the first position in the sort is taken as the reference displacement value. The reference displacement value is then divided by the foot displacement values ​​corresponding to the left and right feet respectively to form the contact coefficient of the left foot and the contact coefficient of the right foot. Thus, the contact coefficient of the foot with the smaller foot displacement value is closer to one, and the contact coefficient of the foot with the larger foot displacement value is closer to zero. This converts the degree of movement of the two feet between two adjacent frames into the degree of load participation of the two feet in the current frame. After determining the contact coefficients for both feet, the positions of the toe points, heel points, and outer edge points of each foot are scaled using the ankle points as base points. Specifically, for any point in one foot, the line connecting that point to the corresponding ankle point is first determined, keeping the direction of the line unchanged. Then, the length of the line connecting that point to the ankle point is multiplied by the contact coefficient of that foot to obtain the scaled length of the line. Finally, the point is repositioned along the original line direction using the scaled length of the line, serving as the effective boundary point corresponding to that point in the current frame. After performing the above processing on the toe points, heel points, and five outer edge points of each foot, a set of effective boundary points corresponding to both feet in the current frame is formed. When the contact coefficient of a foot is equal to one, the points corresponding to that foot remain in their original positions. When the contact coefficient of a foot is between zero and one, the points corresponding to that foot shrink towards the ankle point along the direction from the ankle point to the boundary point. In this embodiment, the ankle point always serves as the scaling center and does not move, thereby ensuring that the set of effective boundary points still corresponds to the actual positions of both feet in the current frame. After the effective boundary point set is formed, each point in the effective boundary point set is projected onto the ground plane, and the x-coordinate of each point is read in the two-dimensional coordinate system corresponding to the ground plane. The projected points are then arranged in ascending order of x-coordinate. After sorting, convex hull calculation is performed. The specific process is as follows: each sorted point is read sequentially, and points are connected one by one to form a boundary candidate chain. The turning value is calculated for three adjacent points in the boundary candidate chain. The turning value is formed by the component of the cross product of the previous side segment vector and the next side segment vector in the normal direction of the ground plane. When the point corresponding to the turning value is located inside the boundary, the middle point among the three points is deleted. When the point corresponding to the turning value is located outside the boundary, the middle point is retained. The above rules are repeated until all remaining points are peripheral points. Then, the peripheral points are connected end to end to form the dynamic support boundary of the current frame. Afterward, the dynamic support boundaries of each frame are collected in chronological order to form a dynamic support boundary sequence. Each frame's dynamic support boundary is a closed boundary, which is subsequently used for the calculation of the directed distance between the centroid projection point and the dynamic support boundary. Through the above processing, the positional changes of the two feet between two adjacent frames can be converted into the contact coefficients of the two feet, and the contact coefficients can be further written into the scaling process of the boundary points of the two feet, thereby forming an effective set of boundary points and a dynamic support boundary sequence corresponding to the current frame's bearing state. This avoids directly writing all the boundary points of the two feet into the support boundary with equal weight, and also avoids the swing foot boundary points from over-expanding the dynamic support boundary, so that the subsequent distance calculation between the human body's centroid projection and the dynamic support boundary is closer to the actual foot bearing range of the current frame. In practical applications: When a subject performs a starting, stepping, or turning action, the system first reads the toe points, heel points, and outer edge points of both feet in two adjacent frames, calculates the distance between each point on the ground plane, and then accumulates them to form the foot displacement values ​​of the left and right feet, thereby obtaining the contact coefficients of the left and right feet. Subsequently, using the corresponding ankle points of each foot as the base point, the toe points, heel points, and outer edge points are scaled according to the contact coefficients to form the effective boundary point set for the current frame. Finally, the effective boundary point set is projected onto the ground plane, and the outer points are retained and the inner points are deleted through convex hull calculation to form the dynamic support boundary corresponding to the current frame. The dynamic support boundary sequence is then collected in chronological order for direct use in subsequent fall risk value calculations.

[0022] S5. Based on the centroid projection sequence and the dynamic support boundary sequence, calculate the directed distance from the centroid projection point to each segment of the dynamic support boundary frame by frame. Sort the directed distances corresponding to each segment in ascending order of absolute value and take the first directed distance as the boundary distance of the current frame. Then, accumulate the boundary distances of each frame with positive boundary distances in time order to form the fall risk value. In this embodiment, S5 is used to calculate the fall risk value based on the centroid projection sequence and the dynamic support boundary sequence. The process is as follows: First, in each frame, each edge segment is formed by the adjacent boundary points of the dynamic support boundary. Then, the directed distance between the centroid projection point and each edge segment is calculated edge by edge. Next, the boundary distance closest to the dynamic support boundary is determined from all directed distances in the current frame, and a boundary distance sequence is formed according to the frame order. Finally, the boundary distances of each frame with positive values ​​in the boundary distance sequence are accumulated in time order to obtain the fall risk value. The positive or negative sign of the directed distance is used to distinguish whether the centroid projection point is located inside or outside the dynamic support boundary, and the boundary distance is used to characterize the degree of the closest boundary crossing of the centroid projection point relative to the dynamic support boundary in the current frame. The implementation process includes the following steps: First, for each centroid projection point in the centroid projection sequence, the dynamic support boundary of the corresponding frame in the dynamic support boundary sequence is read, and adjacent boundary points in the dynamic support boundary are connected sequentially to form each edge segment. The first and last boundary points are also connected to form a closed edge segment. After each edge segment is formed, for the centroid projection point of the current frame, the perpendicular position of the centroid projection point on the line containing each edge segment is calculated. When the perpendicular is between the two endpoints of the corresponding edge segment, the distance from the centroid projection point to the perpendicular is taken as the distance value of the edge segment. When the perpendicular is outside the two endpoints of the corresponding edge segment, the distance from the centroid projection point to the edge segment is calculated. The distance between the two endpoints is calculated, and the shorter distance is taken as the distance value of the edge segment. After the distance value is determined, it is then determined whether the centroid projection point is located inside or outside the area enclosed by the dynamic support boundary of the current frame. The distance value corresponding to the inner side is recorded as a negative value, and the distance value corresponding to the outer side is recorded as a positive value. This forms the directed distance corresponding to each edge segment of the current frame. Among them, when the centroid projection point is located within the area enclosed by the dynamic support boundary, it is determined to be located inside; when the centroid projection point is located outside the area enclosed by the dynamic support boundary, it is determined to be located outside; when the centroid projection point is located on the dynamic support boundary, the corresponding directed distance is recorded as zero. After the directed distances corresponding to each edge segment are formed, all directed distances in the current frame are sorted in ascending order of absolute value, and the directed distance at the top of the sort is taken as the boundary distance of the current frame. The reason for using the top of the sort after absolute value is that the closest position relationship between the centroid projection point and the dynamic support boundary corresponding to this directed distance can directly characterize the minimum out-of-bounds distance or minimum inner distance between the centroid projection point and the dynamic support boundary in the current frame. After the boundary distance of the current frame is determined, the centroid projection points and dynamic support boundaries of each frame are read in chronological order, and the directed distance calculation and boundary distance determination process is repeated. The boundary distances of each frame are collected in frame order to form a boundary distance sequence. Each boundary distance in the boundary distance sequence corresponds one-to-one with the corresponding frame, and all boundary distances are obtained based on the same ground plane and the same dynamic support boundary determination rule. After the boundary distance sequence is formed, the boundary distances of each frame in the sequence are read sequentially in time, and the boundary distances of each frame with a positive boundary distance are selected. Since a positive boundary distance indicates that the centroid projection point of the current frame is located outside the dynamic support boundary, all positive boundary distances are accumulated sequentially to obtain the fall risk value. In this embodiment, a negative boundary distance indicates that the centroid projection point of the current frame is located inside the dynamic support boundary, and a zero boundary distance indicates that the centroid projection point of the current frame falls on the dynamic support boundary. Neither of these is written into the fall risk value accumulation process. Thus, the fall risk value is only composed of the boundary distances of each frame where the centroid projection point exceeds the dynamic support boundary, and it increases with the increase of the number of frames exceeding the boundary and the degree of exceeding the boundary. Through the above processing, the spatial relationship between the centroid projection sequence and the dynamic support boundary sequence can be converted into directed distance frame by frame. Furthermore, the current frame boundary distance can be extracted from the directed distance of each frame, and the positive boundary distances corresponding to the boundary violations can be accumulated in time order to form the fall risk value. This can not only maintain the direct correspondence between the fall risk value and the positional relationship between the current frame centroid projection point and the dynamic support boundary, but also avoid writing stable frames inside the dynamic support boundary into the fall risk value, so that the final fall risk value can more accurately reflect the degree of boundary violation of the subject throughout the entire assessment process. In practical applications: After the subject completes the actions of standing, starting, stepping, or turning, the system first reads the centroid projection points of each frame and the corresponding dynamic support boundary. It then connects the adjacent boundary points of the dynamic support boundary in sequence to form each side segment. Next, it calculates the directed distance from the centroid projection point to each side segment and takes the first directed distance in ascending order of absolute value as the current frame boundary distance. After that, it collects the boundary distances in the frame order to form a boundary distance sequence, and adds up the boundary distances of each frame with positive values ​​in the time sequence to finally obtain the fall risk value corresponding to this assessment process, which is used to output the fall risk assessment results for the elderly.

[0023] Furthermore, a depth camera-based fall risk assessment system for the elderly includes: The image point acquisition module is used to acquire depth image sequences corresponding to the subject's assessment actions, and to identify the three-dimensional coordinates of human bone points and the toe points, heel points, ankle points and knee points of both feet frame by frame to form a bone point sequence; The centroid determination module divides the human body into segments based on the bone point sequence, calculates the centroid coordinates of the human body according to the segment mass value and centroid position of each segment, and projects the centroid coordinates of the human body onto the ground plane to form a centroid projection sequence. The outer edge positioning module constructs the foot axis based on the toe and heel points of each foot in the bone point sequence, constructs the lower leg axis based on the corresponding knee and ankle points, divides the foot axis into six equal parts to obtain five sub-points, determines the lateral direction of the foot by the cross product direction of the lower leg axis and the foot axis, and determines five outer edge points of the foot based on the length correction amount corresponding to each of the five sub-points, thus forming the outer edge sequence of the foot. The boundary generation module calculates the contact coefficients of the two feet based on the position changes of the two feet between adjacent frames in the bone point sequence, scales the positions of the toe points, heel points and outer edge points of the two feet according to the contact coefficients to form an effective boundary point set, and performs convex hull calculation on the effective boundary point set to form a dynamic support boundary sequence. The risk calculation module calculates the directed distance from the centroid projection point to each segment of the dynamic support boundary frame by frame based on the centroid projection sequence and the dynamic support boundary sequence. The directed distances corresponding to each segment are sorted in ascending order of absolute value, and the directed distance at the top of the sort is taken as the boundary distance of the current frame. The boundary distances of each frame with a positive boundary distance are then accumulated in time order to form the fall risk value.

[0024] Working principle: This scheme first uses a depth camera to continuously acquire depth images of the subject, extracting bone points such as toe, heel, ankle, and knee points, and calculating the coordinates and projection position of the human centroid in each frame. Then, based on the foot axis, lower leg axis, and the changes between consecutive frames, combined with bone point temporal relationship pattern recognition and lateral determination, the outer edge of the foot is completed, obtaining a foot outer edge point in each frame that more closely approximates the actual foot bearing capacity. On this basis, the contact coefficient of the feet is determined based on the positional changes of the feet between adjacent frames, and the foot boundary points are scaled to generate a dynamic support boundary. Finally, the human centroid projection is compared frame by frame with the dynamic support boundary to calculate the degree to which the centroid projection exceeds the support boundary, and the excess portion is accumulated in chronological order to form a fall risk value, thus completing the fall risk assessment for the elderly. The entire process is a series of steps: bone point extraction, centroid calculation, foot outer edge reconstruction, support boundary generation, and risk value calculation. In practical applications, such as hospital outpatient screening, rehabilitation reassessment, or daily assessment in elderly care facilities, subjects only need to stand in front of the depth camera and perform actions such as standing, starting, stepping, or turning. The system can automatically acquire changes in bone points and foot boundaries during the movement and simultaneously display the position of the human body's center of mass, the support range of the feet, and the risk changes in the virtual reality interface. If the subject takes too large a step or turns unsteadily at a certain moment, causing the projection of the human body's center of mass to gradually approach or even exceed the current foot support range, the system will reflect the out-of-bounds distance on the dynamic support boundary at that moment and continuously accumulate to form a higher fall risk value. In this way, the balance judgment that originally relied on manual observation is transformed into a quantitative result that can be directly displayed and continuously compared.

[0025] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assessing fall risk in the elderly based on a depth camera, characterized in that, include: S1. Obtain the depth image sequence corresponding to the subject's assessment action, and identify the three-dimensional coordinates of human bone points and the toe points, heel points, ankle points and knee points of both feet frame by frame to form a bone point sequence. S2. Divide the human body into segments according to the bone point sequence, calculate the human body centroid coordinates according to the segment mass value and centroid position of each segment, and project the human body centroid coordinates onto the ground plane to form a centroid projection sequence. S3. Construct the foot axis based on the toe and heel points of each foot in the bone point sequence, construct the lower leg axis based on the corresponding knee and ankle points, divide the foot axis into six equal parts to obtain five points, determine the lateral direction of the foot by the cross product direction of the lower leg axis and the foot axis, and determine five outer edge points of the foot based on the length correction amount corresponding to each of the five points to form the outer edge sequence of the foot. S4. Calculate the contact coefficients of the two feet based on the position changes of the two feet between adjacent frames in the bone point sequence. Scaling the positions of the toe points, heel points and outer edge points of the two feet according to the contact coefficients to form an effective boundary point set. Perform convex hull calculation on the effective boundary point set to form a dynamic support boundary sequence. S5. Based on the centroid projection sequence and the dynamic support boundary sequence, calculate the directed distance from the centroid projection point to each segment of the dynamic support boundary frame by frame. Sort the directed distances corresponding to each segment in ascending order of absolute value and take the first directed distance as the boundary distance of the current frame. Then, accumulate the boundary distances of each frame with positive boundary distances in chronological order to form the fall risk value.

2. The method for assessing fall risk in the elderly based on a depth camera according to claim 1, characterized in that: S1 includes: S1-1. Convert each frame of depth image in the depth image sequence into a three-dimensional coordinate point set, extract the human body connected region and human body contour, arrange the height coordinates of each point of the human body contour in ascending order, take the contour point corresponding to the first position of the sort to form the foot point set, and then separate the foot point set into the left foot point set and the right foot point set according to the human body midline to form the foot point set sequence. S1-2. For each foot in the foot point set sequence, calculate the distance between each pair of points in the foot point set, arrange the distances between the points in ascending order, take the pair of points corresponding to the last position in the sort to form the foot axis, determine the end closer to the human body's central axis as the heel point, determine the end farther from the human body's central axis as the toe point, and then determine the intersection of the midpoint of the foot axis along the height direction with the human body's central axis as the ankle point, thus forming the foot point sequence. S1-3. Extract the center line from the connected regions of the human body in each frame, determine the lower limb inflection point above the ankle point in the center line as the knee point, and write the bifurcation point and endpoint of the center line into the human bone points to form a bone point sequence.

3. The method for assessing fall risk in the elderly based on a depth camera according to claim 2, characterized in that: S2 includes: S2-1. Divide the body into trunk, thigh, calf and foot segments according to the connection relationship between adjacent bone points in the bone point sequence. Select multiple sampling points at equal intervals along the segment axis corresponding to each human body segment to form a sampling point sequence. S2-2. For each sampling point in the sampling point sequence, find the intersection with the human body contour along the vertical direction of the segment axis to obtain the left and right intersection points. Calculate the distance between the left and right intersection points, and multiply the distance corresponding to each sampling point with the position distance of the sampling point on the segment axis. Then, accumulate the results according to the human body segments to form a segmented quality value sequence.

4. The method for assessing fall risk in the elderly based on a depth camera according to claim 3, characterized in that: S2 further includes: S2-3. For each sampling point in the sampling point sequence, multiply the position distance of each sampling point on the segment axis by the corresponding distance and accumulate them according to the human body segment. Then divide the accumulated result by the accumulated distance value corresponding to the human body segment to form the segment centroid position sequence. S2-4. Determine the centroid coordinates of each human body segment based on the segmented mass value sequence and the segmented centroid position sequence. Then, weight and accumulate the centroid coordinates of each human body segment according to the segmented mass value sequence to obtain the centroid coordinates of the human body. Finally, project the centroid coordinates of the human body onto the ground plane to form a centroid projection sequence.

5. The method for assessing fall risk in the elderly based on a depth camera according to claim 4, characterized in that: S3 includes: S3-1. Construct the foot axis based on the toe and heel points of each foot in each frame of the bone point sequence, and construct the lower leg axis based on the corresponding knee and ankle points. Divide the foot axis into six equal parts to obtain five sub-points, and calculate the projection value of the cross product of the lower leg axis and the foot axis onto the normal direction of the ground plane to form a direction value sequence. S3-2. For the current frame in the direction value sequence, extract the direction value of the previous frame, the direction value of the current frame, and the direction value of the next frame. Compare the signs of the direction values ​​of the previous frame and the current frame, and the direction values ​​of the current frame and the next frame. Calculate the absolute differences between the direction values ​​of the current frame and the previous frame, and the absolute differences between the direction values ​​of the current frame and the next frame. If the two sign comparison results are consistent and the ranking positions corresponding to the differences of the two absolute values ​​are both at the first position, the side corresponding to the current frame direction value is determined as the outer side. If the two sign comparison results are inconsistent and the signs of the direction values ​​of the previous frame and the next frame are consistent, the two sides of the foot axis are determined as the first candidate side and the second candidate side, respectively. If the two sign comparison results are consistent and the ranking positions corresponding to the differences of the two absolute values ​​are not both at the first position, the side corresponding to the direction value of the previous frame and the side corresponding to the direction value of the next frame are determined as the side to be detected.

6. The method for assessing fall risk in the elderly based on a depth camera according to claim 5, characterized in that: S3 further includes: S3-3. For the first candidate side, the second candidate side, or the side to be tested, retrieve the first confirmed frame in chronological order where the outer edge of the foot has been formed, and retrieve the first confirmed frame in chronological order where the outer edge of the foot has been formed. On each side, determine five base points at five points according to 50 mm for males and 35 mm for females. Calculate the cumulative distance between the five base points corresponding to each side and the corresponding outer edge of the foot in the previous confirmed frame, the cumulative distance between the five base points corresponding to each side and the corresponding outer edge of the foot in the subsequent confirmed frame, and the difference between the cumulative distance between the previous and subsequent confirmed frames. Then, sort the sum of the cumulative distance between the previous and subsequent confirmed frames and the difference between the cumulative distance between the previous and subsequent confirmed frames in ascending order. Take the side that is first in both sorts and determine it as the outer side.

7. The method for assessing fall risk in the elderly based on a depth camera according to claim 6, characterized in that: S3 further includes: S3-4. Determine the lateral direction of the foot based on the lateral side, calculate the angle between the lower leg axis and the foot axis in the current frame, the angle between the lower leg axis and the foot axis in the previous confirmation frame, and the angle between the lower leg axis and the foot axis in the subsequent confirmation frame. Distribute the difference between the average of the angle in the current frame, the angle in the previous confirmation frame, and the angle in the subsequent confirmation frame to five points according to the ratio of the distance from the heel point to the length of the foot axis, forming a length correction sequence. S3-5. Along the outer side of the foot, locate the five points as five outer edge points of the foot according to the sum of the length correction amount corresponding to the basic length and the length correction amount sequence, and collect the five outer edge points of the foot in time order to form the outer edge sequence of the foot.

8. The method for assessing fall risk in the elderly based on a depth camera according to claim 7, characterized in that: S4 includes: S4-1. For the toe point, heel point and outer edge point of the feet in two adjacent frames in the bone point sequence, calculate the distance between each point in the ground plane, and accumulate the distance of all points corresponding to each foot to form the foot displacement value. Then, sort the foot displacement values ​​corresponding to both feet in ascending order, take the first foot displacement value in the sorted order and divide it one by one with the foot displacement values ​​corresponding to each foot to form the contact coefficient corresponding to both feet. S4-2. Using the ankle points of each foot as the base points, keep the direction of the lines connecting the toe points, heel points, and outer edge points of each foot to the ankle points unchanged, and scale the length of each line according to the contact coefficient to form an effective set of boundary points. S4-3. Project the effective boundary point set onto the ground plane, arrange the projected points in ascending order of their horizontal coordinates and perform convex hull calculation. The convex hull calculation is as follows: calculate the turning values ​​of three adjacent points in sequence, delete the middle point located on the inner side, retain the outer points and connect them end to end to form a dynamic support boundary sequence.

9. A method for assessing fall risk in the elderly based on a depth camera according to claim 8, characterized in that: S5 includes: S5-1. For each centroid projection point in the centroid projection sequence, connect adjacent boundary points in the dynamic support boundary sequence to form each edge segment. Calculate the position of the perpendicular foot of the centroid projection point on the straight line of each edge segment. When the perpendicular foot is inside the edge segment, take the distance from the centroid projection point to the perpendicular foot as the directed distance of the edge segment. When the perpendicular foot is outside the edge segment, take the distance from the centroid projection point to the shortest endpoint among the edge segment endpoints as the directed distance of the edge segment. Record the directed distance corresponding to the centroid projection point being inside the dynamic support boundary as a negative value and the directed distance corresponding to the centroid projection point being outside the dynamic support boundary as a positive value. S5-2. Sort the directed distances corresponding to each edge segment in ascending order of absolute value, take the first directed distance in the sorted order as the current frame boundary distance, and collect the current frame boundary distances in frame order to form a boundary distance sequence. S5-3. The boundary distances of each frame with a positive boundary distance in the boundary distance sequence are accumulated in chronological order to form a fall risk value.

10. A fall risk assessment system for the elderly based on a depth camera, characterized in that, include: The image point acquisition module is used to acquire depth image sequences corresponding to the subject's assessment actions, and to identify the three-dimensional coordinates of human bone points and the toe points, heel points, ankle points and knee points of both feet frame by frame to form a bone point sequence; The centroid determination module divides the human body into segments based on the bone point sequence, calculates the centroid coordinates of the human body according to the segment mass value and centroid position of each segment, and projects the centroid coordinates of the human body onto the ground plane to form a centroid projection sequence. The outer edge positioning module constructs the foot axis based on the toe and heel points of each foot in the bone point sequence, constructs the lower leg axis based on the corresponding knee and ankle points, divides the foot axis into six equal parts to obtain five sub-points, determines the lateral direction of the foot by the cross product direction of the lower leg axis and the foot axis, and determines five outer edge points of the foot based on the length correction amount corresponding to each of the five sub-points, thus forming the outer edge sequence of the foot. The boundary generation module calculates the contact coefficients of the two feet based on the position changes of the two feet between adjacent frames in the bone point sequence, scales the positions of the toe points, heel points and outer edge points of the two feet according to the contact coefficients to form an effective boundary point set, and performs convex hull calculation on the effective boundary point set to form a dynamic support boundary sequence. The risk calculation module calculates the directed distance from the centroid projection point to each segment of the dynamic support boundary frame by frame based on the centroid projection sequence and the dynamic support boundary sequence. The directed distances corresponding to each segment are sorted in ascending order of absolute value, and the directed distance at the top of the sort is taken as the boundary distance of the current frame. The boundary distances of each frame with a positive boundary distance are then accumulated in time order to form the fall risk value.

Images