CN102551731B - Tumbling movement detecting method based on data curve comparison - Google Patents

Abstract

A fall activity detection method based on data curve comparison belongs to the innovative application of pervasive computer human-computer interaction technology. The falling activity detection method introduces changes in the surrounding magnetic field to reflect the movement of body parts in human activities, grasps the complex behavioral feature information in human activities, and uses the self-designed curve shape context investigation and curve feature matrix to obtain the perception data curve. The curve comparison technical scheme of generation and difference degree calculation realizes the detection of falling activities, which fully meets the real-time and high-efficiency requirements of activity detection in human-computer interaction in pervasive computing applications.

Description

A Fall Activity Detection Method Based on Data Curve Comparison

technical field

The invention belongs to the field of pervasive computing application technology, and relates to a human body activity detection method in human-computer interaction technology. The method effectively detects people's fall activity by comparing and processing the perception data curves generated with human body activities. the

Background technique

As a typical application of ubiquitous computing, human fall activity detection is also an important key technology in activity-oriented human-computer interaction. It is of fundamental and important significance for avoiding the serious health hazards caused by falls to people, especially the elderly, and for gradually and deeply studying the accurate detection of various human activity information in human-computer interaction. Fall activity detection technology focuses on converting the qualitative description of the activity into a quantitative description reasonably, so as to realize the identification and judgment of the detection system. The existing fall activity detection research work mainly adopts three types of detection methods: database-based motion classification and recognition detection, acceleration-based detection and image processing-based detection. the

The database-based action classification, recognition and detection method stores the collected behavior data accompanying various human activities into the database and analyzes them to extract different features corresponding to different activities, so as to analyze the user's behavior activities according to these features. Carry out classification and identification to judge whether there is a fall activity. The establishment of human activity perception database provides powerful data support for fall activity detection, but the problem is that the work of collecting enough data and establishing a database for individual users is very complicated and cumbersome, which greatly affects the feasibility of this work . the

Acceleration-based fall activity detection is a more widely used method, which mainly relies on examining acceleration data to set thresholds for detection. It collects acceleration data through the accelerometer sensor, and establishes detection and judgment standards according to different processing methods of data and different setting methods of threshold values: either based on the absolute value of the peak value of acceleration, or based on the total vector sum of acceleration, dynamic vector sum of acceleration, vertical acceleration and Dynamic indicators such as the difference between the maximum and minimum acceleration values are used for judgment. The problem with this type of detection method is that acceleration data changes will occur in many human activities, and the fall activity detection based on this is prone to a large false alarm rate, and the specificity of detection is poor. the

Image processing-based fall activity detection uses a method that captures images of human activity and detects "visible" falls based on image processing techniques. However, the acceptability and affordability of such methods are not good, and due to the use of image technology for body activity detection, the detection area is completely limited to a very limited monitorable range, and the high cost of setting the monitorable range limits the detection area extension. the

Contents of the invention

The purpose of the present invention is to provide a new type of detection method for falling activities in order to detect falling activities more effectively and avoid the health hazards caused by falls, so as to solve the key technical problems of human-computer interaction detection technology and accurately and efficiently detect , Distinguish falling activities from other routine human activities, and improve the "sensitivity" and "specificity" of the detection results. the

In order to realize the purpose of the above invention, the present invention adopts the following technical solutions. The fall activity detection method based on data curve comparison uses a variety of sensors to collect real-time sensory data that is accompanied by human activities, contains accurate details of the activity, and can be used for subsequent data curve processing and comparison, forms signal curves based on the sensory data, and examines the curves The shape feature of the curve is used to detect the fall activity by quantitative comparison of the curves. Different from fall detection methods that rely on acceleration features, this method not only uses accelerometer sensors to record acceleration data accompanying human activities, but also innovatively uses magnetic field sensing sensors to collect data, placing magnetic field sensing sensors and miniature magnetic accessories on effective parts of the body , use the magnetic field sensing sensor to perceive the change of the surrounding magnetic field, and obtain the magnetic field strength data reflecting the characteristics of human body activities, which reflects the position and distance changes between the body parts where the two devices (sensors and magnetic accessories) are placed during activities. the

The fall activity detection method based on data curve comparison includes a set of activity detection framework, as shown in Figure 1, which is divided into three parts: generation of human activity characteristic perception data curve, perception data curve shape context inspection and perception data curve similarity comparison. Corresponding to several steps of the fall activity detection method, as follows:

Step 1: Generate a perception data curve based on real-time data collected by various sensors, such as acceleration data, direction data, and magnetic field strength data. the

Step 2: Generate a shape context histogram from the perceptual data curve. Specifically, a feature point S is selected on the perceptual data curve, and a logarithmic polar coordinate system covering all other points on the curve is generated centering on the selected feature point. R is the maximum distance from the center point of the coordinate system to other points. The coordinate system is evenly divided into kd distance ranges according to the logR value, denoted as D1, D2, ..., Dkd, and the polar coordinate central angle 2π is equally divided into ka angle ranges, denoted as A1, A2, ... Aka. The logarithmic polar coordinate system is divided into kd×ka histogram intervals covering the entire curve with the feature point S as the center, as shown in Figure 2. Any point other than the feature point S will be within a certain distance and angle range of the coordinate system according to its polar coordinate value (ρ, θ), that is, it will fall within a certain histogram interval centered on the feature point S, The shape context of point S on the curve can be obtained by counting the number of points in each histogram interval. the

Step 3: Convert the shape context histogram to a matrix representation, i.e. generate a curve feature matrix. Define the shape context matrix Ms of the feature point S, which corresponds to the number of histogram intervals. The matrix Ms has kd×ka elements, where the value of the element (i, j) falls in the coordinate system. The distance range is Di, and the angle range is The number of points in the interval of Aj. Similarly, the shape context histograms of other feature points X are also transformed into their shape context matrix MX. Because the shape context of a feature point embodies the characteristics of the part of the curve centered on this point, the combination of the shape context matrix of multiple feature points can reflect the characteristics of the entire curve. The shape context matrix of all feature points is spliced and combined to generate a matrix M' of nkd×ka, M' is the feature matrix of the perceptual data curve, where n is the number of selected feature points. The curve feature matrix can be visually represented in the form of a grayscale image, as shown in Figure 3, the darker the color in the matrix box, the greater the value of the corresponding element. The image representations of the curve feature matrix for the same kind of activities are similar, while the images corresponding to different activities will be quite different. the

Step 4: After obtaining the characteristic matrix of the curve, the comparison of the curves can be transformed into the comparison of its characteristic matrix. The similarity comparison of the curve feature matrix can be quantitatively calculated by the Hausdorff distance measurement method of its image representation. The size of the comparison results indicates the degree of approximation of the data curves corresponding to each feature matrix—the smaller the value, the higher the degree of approximation. This comparative value is also the output result of quantitatively comparing the sensory data curve and determining the falling activity. For a set of sensory data corresponding to unknown activities, it is compared with another set of sample sensory data corresponding to known falling activities. If two If the value of the comparison result between the sensory data curves is small enough, it is determined that the unknown activity is a fall activity. the

The beneficial effect of the present invention is that the falling activity detection method combines the research on the characteristics of body activities, novelly introduces changes in the surrounding magnetic field to reflect the movement of body parts in human activities, grasps the complex behavioral feature information in human activities, and analyzes the obtained perception The data curve realizes the real-time and efficient detection of falling activities through the self-designed curve comparison technology scheme based on the curve shape context investigation, curve feature matrix generation and difference calculation, which overcomes the inability to simply use the acceleration threshold condition to determine the falling activity. It fully reflects the characteristic difference between falling activity and normal human activity, and it is difficult to reconcile the detection effect "sensitivity" and "specificity", which meets the needs of falling activity detection in the practical application of ubiquitous computing. the

Description of drawings

Fig. 1 is a schematic diagram of a detection framework of a fall activity detection method based on data curve comparison. the

Fig. 2 is a schematic diagram of the histogram interval of the logarithmic polar coordinates of the curve shape context. the

Figure 3 is a grayscale image representation of the curve feature matrix. the

Detailed ways

In this fall activity detection method, the sensing data of the magnetic field sensing sensor represents the magnetic field strength around it, and the relative position between the sensor and the magnetic accessory can be deduced by using this data value. By setting the position of the sensor and the magnetic accessory reasonably, it can effectively To accurately capture the common features of body part motion in fall activities to achieve fall detection. Specifically, the magnetic attachment can be placed on the leg, slightly above the knee, while the sensor is placed near the pant pocket on the other leg. The magnetic field intensity perception data collected by the sensor characterizes the movement between the legs during human activities. the

After obtaining the sensory data including human activity characteristics and generating data curves, the fall activity detection method based on data curve comparison is to examine the shape context of the sensory data curves. The process includes the selection of shape context feature points—determining the number and location of feature points, And the division of the interval of the logarithmic polar coordinate system. Since the posture of the legs often changes significantly during the fall process, it will cause regular changes in the magnetic field strength and sensory data. These changes should be considered when selecting feature points in the data curve shape context investigation. In this fall activity detection method, two "end points of decline" in a pair of "continuous decline" are selected as the characteristic points of the data curve. the

"Continuous decline" means: if the sensory data curve shows that the reading has a monotonous decline in the time period [t1, t2], followed by a monotonous rise in the time period [t2, t3], and then in [t3, The monotonous decline again in the time period of t4] means that the two declines occurring in the time period [t1, t2] and [t3, t4] are continuous, that is, a pair of "continuous declines", and the lowest of the two declines in the continuous decline is The points are called "descent endpoints", respectively. the

The logarithmic polar coordinate system is generated with the characteristic point of the data curve as the center point. In this coordinate system, 5 distance ranges are evenly divided according to the logarithmic value logR of the maximum distance between other nodes and the center point, which are recorded as D1, D2,..., D5, at the same time divide the central angle 2π into 12 angle ranges, denoted as A1, A2, ..., A12. the

In the implementation of the fall activity detection method based on data curve comparison, first calculate the cumulative sum of the two drops in each pair of "continuous drops" in the perceived data curve, and check whether it exceeds the threshold Thtm, if the check result is true, then in the second time Immediately execute the follow-up operation within 1.6 seconds after the "drop end point" of the drop; the follow-up operation is to use the two drop end points in the current "continuous drop" as feature points, calculate the shape context of the curve, and combine them to generate a curve feature matrix, specifically The two feature points are respectively used as the center point of the logarithmic polar coordinate system, and any other point on the data curve falls into a specific distance range and angle range according to its polar coordinate value (represented by the distance ρ and the angle θ), That is, within an interval of the logarithmic polar coordinate system, a 5×12 shape context matrix MX is obtained, and the value at the position (i, j) in the matrix represents the number of points falling into the distance range Di and the angle range Aj . Combining the shape context matrices generated by the two feature points respectively, a 10×12 curve feature matrix M is obtained. Then the curve feature matrix M is represented by a grayscale image, and the Hausdorff distance between it and the known feature matrix of the corresponding fall activity is calculated, and it is checked whether it is smaller than the threshold Thhd. If the result is true, the current perception data curve corresponds to activity is marked as "fall", otherwise marked as "non-fall". the

Claims (2)

1. A fall activity detection method based on data curve comparison, is characterized in that, comprises the following steps: 1) Process the sensory data collected by sensors along with human activities, and generate sensory data curves based on real-time data collected by various sensors, including acceleration data, direction data, and magnetic field strength data; 2) Use the shape context to examine the perception data curve, and generate a shape context histogram based on the perception data curve: select a feature point S on the perception data curve, and generate a logarithmic pole covering all other points on the curve with the selected feature point as the center Coordinate system, R is the maximum distance from the center point of the coordinate system to other points, the coordinate system is evenly divided into kd distance ranges according to the logR value, denoted as D1, D2, ..., Dkd , and the polar coordinate center angle 2π Equally divided into ka angle ranges, denoted as A1, A2, ... , Aka ; the logarithmic polar coordinate system is divided into kd×ka histogram intervals covering the entire curve with the feature point S as the center, except for the feature point Any point other than S will be within a certain distance and angle range of the coordinate system according to its polar coordinate value (ρ, θ), that is, within a certain histogram interval centered on the feature point S, and each The number of points in the histogram interval can get the shape context of the feature point S on the curve; The feature points are two "end points of decline" in a pair of "continuous decline". There is a monotonous rise in the time period [t2,t3], and then a monotonous decline again in the time period [t3,t4], it is said that the two declines occurred in the time period [t1,t2] and [t3,t4] It is continuous, that is, a pair of "continuous declines", and the lowest points of the two declines in the continuous decline are respectively called the "end of decline"; 3) Convert the shape context histogram into a matrix representation, that is, generate a curve feature matrix; define the shape context matrix Ms of the feature point S , corresponding to the number of histogram intervals, the matrix Ms has kd×ka elements, and the elements (i , The value of j ) is the number of points falling in the range of distance Di and angle range Aj in the coordinate system; similarly, the shape context histogram of other feature points X is also converted into its shape context matrix MX ; because feature points The shape context of , embodies the characteristics of the curve part centered on this point, so combining the shape context matrix of multiple feature points can reflect the characteristics of the entire curve; the shape context matrix of all feature points is spliced and combined to generate an nkd The matrix M′ of × ka, M′ is the feature matrix of the perceptual data curve, where n is the number of selected feature points; 4) After obtaining the characteristic matrix of the curve, the comparison of the curve is transformed into the comparison of its characteristic matrix, and the comparison calculation of the difference degree of the characteristic matrix image is carried out by using the Hausdorff distance, which is realized by comparing with the characteristic matrix image of the sensory data curve of the known fall activity Real-time efficient detection of fall activity. 2. A method for detecting fall activity based on data curve comparison according to claim 1, characterized in that: the change in the surrounding magnetic field strength is used to reflect the change in the distance between the magnetic field induction sensor and the magnetic attachment, thereby reflecting the body parts in human activities By reasonably setting the positions of sensors and magnetic accessories and collecting magnetic field strength perception data, the complex behavioral feature information in falling activities can be grasped.

Images