A real-time fall event detection method based on surveillance videos

Abstract

The present invention relates to the field of image recognition, and specifically discloses a real-time fall event detection method based on surveillance videos, including: S1 image acquisition step, real-time acquisition of video image information of a target area, and the pixel value C of the point with coordinates (x, y) in the k-th frame of the video image information k (x, y) is subtracted from the pixel value Bk(x, y) of the corresponding point in the background image to obtain a difference image D k (x, y), and the difference image D k (x, y) is compared with a set threshold T to obtain a binary image R k (x, y); S2 human body posture recognition step, the binary image R k (x, y) is used with an RVM classifier to determine the target posture category, and then the target posture category is input into an HMM model as a target posture value sequence for evaluation. The target posture categories include: standing, tilting, lying flat, and others. It can accurately recognize the high-frequency behaviors of users and avoid false alarms and missed identifications.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and particularly to a method for real-time fall event detection based on surveillance videos. Background Art

[0002] In recent years, smart home services provided for the elderly living alone have become a popular research field. According to surveys, among all the dangerous events that occur to the elderly living alone, fall events pose the greatest harm, which may cause the fallen person to lose consciousness, become disabled or even die. Therefore, it is very necessary to detect possible fall behaviors in a timely manner and alert family members so that the elderly can receive assistance.

[0003] Currently, human fall detection technologies mainly include: fall detection technologies based on video and sound, and fall detection technologies based on pressure. Among them, the fall detection technology based on video and sound uses cameras and other sound collection tools to monitor a certain area, and then detects human fall events in the monitored images through human recognition and pose recognition in image processing; the fall detection technology based on pressure uses pressure sensors laid on the floor to determine events such as impacts through the pressure changes of objects on the ground, thereby detecting falls.

[0004] Due to the large-scale use of video surveillance devices, the fall detection technology based on video (also known as computer vision technology) has become a popular technology. In the surveillance video detection method, most studies only use a single motion feature or pose feature, so it is easy to cause a large number of false detections. On the other hand, relevant literature does not consider dealing with fall events along the direction of the camera's illumination. In such cases, the shape of the fallen person is similar to that of a standing person, and it is very difficult to distinguish the two postures based solely on appearance features.

[0005] Therefore, in the literature with Chinese Patent Publication No. CN 103955699 A, a method for real-time fall event detection based on surveillance videos installs multiple cameras with different shooting angles facing the same target area in the detection scenario. The multiple cameras continuously shoot the target area, including the following steps: multiple cameras simultaneously shoot a segment of video of the target area; from the multiple videos of the same time period respectively shot by the multiple cameras, the foreground images representing the target are extracted from each frame of the picture; the respective shape and position features of the foreground images of the same target at the same moment in the pictures shot by the multiple cameras are extracted, and an RVM classifier is used to determine the target pose category corresponding to each frame of the picture; the target pose categories of each frame of the picture obtained are used as a target pose value sequence and input into an HMM evaluator to obtain the posterior probability of the change in the target pose category. The change in the target pose category represents the occurrence of a target fall event; if the posterior probability is greater than a predetermined threshold, it is determined that a fall has occurred.

[0006] The above - mentioned solution mainly determines whether a fall occurs by recognizing the human body's posture and the change speed of the human body's center of gravity through images. However, in daily life, the recognition accuracy has always been unacceptable to users. From the perspective of posture recognition, normal daily postures of users such as sitting upright, sitting and lying down, and lying down are prone to false alarms. Even when adding the recognition of the change speed of the human body's center of gravity, false alarms are still likely to be triggered due to individual differences (the difference in the speed of some individuals' movements, resulting in the setting of the threshold value from the change rate of the center of gravity, which is easy to trigger false alarms). Especially, the situation of falling is relatively complex, including falling during walking (at this time, the center of gravity drops relatively fast, and the scene is also relatively complex), fainting during normal behavior (at this time, the change speed of the center of gravity is uncertain, it may be fast or slow, and the scene is also relatively complex), combined with the user's own daily actions (including sitting down, sitting and lying down, and lying down). These are high - frequency behaviors. If accurate recognition is not carried out, false alarms or missed recognition are very likely to occur, seriously affecting the user experience. Summary of the Invention

[0007] The present invention provides a real - time fall event detection method based on a surveillance video, which can accurately recognize the high - frequency behaviors of users and avoid false alarms and missed recognition.

[0008] To solve the above - mentioned technical problems, the present application provides the following technical solutions:

[0009] A real - time fall event detection method based on a surveillance video, comprising:

[0010] S1 Image acquisition step: Real - time collect the video image information of the target area, subtract the pixel value C k (x, y) of the point with coordinates (x, y) in the k - th frame of the video image information from the pixel value Bk(x, y) of the corresponding point in the background image to obtain the difference image D k (x, y), and compare the difference image D k (x, y) with the set threshold T to obtain the binary image R k (x, y);

[0011] S2 Human posture recognition step: Use the RVM classifier for the binary image R k (x, y) to determine the target posture category, and then input the target posture category as a target posture value sequence into the HMM model for evaluation to obtain the posterior probability of the change of the target posture category. The target posture categories include: standing, tilting, lying flat, and others;

[0012] S3 Image calibration step: Manually calibrate all rest areas in the video image information to obtain the rest area image Xk(x, y);

[0013] S4 scene recognition step. If the posterior probability is greater than a predetermined threshold, it is determined that the target is suspected of falling, then the flash is activated. According to the brightness change of the video image information, it is determined whether the target head posture has changed. If the target head posture has not changed, it is judged as a fall. If the target head posture has changed, an alternating bright and dark grid light is emitted to the target area, and the video image information of the target area is collected again. The binary image R is determined according to the grid distortion situation. k (x, y) and the rest area image X k Based on the depth relationship between (x, y) and the rest area image X, it is determined whether the target and the rest area are at the same depth. If they are at the same depth and their abscissas coincide, it is judged that the target has not fallen; if the target and the rest area are not at the same depth, it is judged that the target is a secondary suspected fall.

[0014] S5 human posture final determination step. If it is judged that the target is a secondary suspected fall, according to the video image information, the direction in which the target head is facing is determined. If there is a highlighted area in the facing direction, it is judged that the target has not fallen; if the target head is facing the ground or the sky, it is judged that the target has fallen.

[0015] The principle and beneficial effects of the basic solution are as follows: In S1, through a method similar to frame difference method, the difference image is obtained by comparing and subtracting the real-time video image information with the background image (pre-set or an image that has not changed for more than a preset time collected in advance, and the preset time can be a time period or duration). Then, further processing is performed on the difference image to obtain a binary image. The binary image obtained in this way is convenient for subsequent posture processing and recognition.

[0016] After obtaining the binary image, in fact, the contour information of the target is obtained, and the RVM classifier is used to identify the target posture to determine the target posture category. Then it is input into the HMM model for evaluation. In short, in fact, the state information of the target posture is determined by identifying the length-width ratio and bone recognition of the binary image. In fact, the above process already has mature technologies in the prior art and will not be elaborated here. The aforementioned technology is essentially the recognition of the target state. Although the state of the target can be obtained, it is impossible to judge whether it is caused by a fall or the target's own will.

[0017] Therefore, based on the previous solution, this solution adds the calibration of the rest area, which is actually pre-calibrated manually. That is, the user (the guardian of the target) can determine items such as sofas, chairs, beds, etc. as the rest area by clicking. When the target is determined to be a suspected fall, the flash is activated. At this time, if the target has fallen, the target's attention should be immersed in self-examination or cause investigation. At this time, the target's response to external stimuli will be relatively weak, and the stimulation of the flash will not attract the target's attention, and the target's head will basically remain unchanged. Therefore, judging as a fall at this time can be highly reasonable.

[0018] Of course, using whether there is an emergency response to the flash as the sole criterion is still prone to problems. Therefore, even if there is an emergency response, this solution will send alternating bright and dark grid lights to obtain depth information for reconfirmation. By collecting the grid distortion, it can be determined whether the target is located in the rest area. If it is located in the rest area, it is determined that the target actively approaches the rest area and changes its posture to rest, so it is judged as not fallen. If it is not in the rest area, the target may have fallen or there may be other reasons, so further investigation is needed.

[0019] In this solution, for such situations, the head orientation of the target is obtained again. If there is a high-brightness area in the direction the head is facing, it is usually a device such as a mobile phone / TV. The target may change its posture to use the relevant device, so the risk is excluded and it is determined as not fallen. If the target's head is facing the ground (most likely tripped) or facing the sky (most likely fallen or slipped, because the judgment of whether in the rest area has been made before, so it can be determined that the target did not lie down actively), it can be determined that the target has fallen. The situations of false alarms and missed reports are reduced.

[0020] Based on determining that the target is a suspected fall, this solution adds a composite judgment of the rest area, whether there is a high-brightness area in the direction the head is facing, and the final direction the head is facing, which can accurately exclude high-frequency behaviors of the target such as watching TV, playing with mobile phones, sleeping, etc., and avoid false alarms and missed identifications.

[0021] Furthermore, the RVM classifier includes 3 binary classifiers. Among them, the first binary classifier distinguishes the first three of the four categories from the fourth category, the second binary classifier distinguishes the first two of the four categories from the third category, and the third binary classifier distinguishes the second category from the first category of the four categories.

[0022] Furthermore, the training process of the RVM classifier includes the following steps:

[0023] Step 11: Extract the shape and position features from the training sample videos belonging to the four categories respectively to form a feature matrix X, and the target pose categories corresponding to all the training sample videos form a vector Y;

[0024] Step 12: Use the RBF kernel function to map the feature matrix X to a high-dimensional space;

[0025] Step 13: According to the Bayesian criterion, solve the optimal weight distribution and distribution parameters as the RVM parameters.

[0026] Further, it also includes S4-1 to identify high-risk areas, and determine the areas with standing postures and rapid changes in the binary image R k (x, y) and the manually calibrated channel areas as high-risk areas. If it is judged that no suspected fall appears in the high-risk areas, it is determined as the target fall.

[0027] When a user shows a suspected fall, if it is in a high-risk area, it is directly judged as a fall, which can improve the recognition efficiency.

[0028] Further, in the S4 scene recognition step, after the flash is turned on, if a highlighted area is captured in the binary image R k (x, y), the target is determined to be a fall.

[0029] When there is water stain on the ground, it is also a cause for the user to fall. At this time, the flash can timely check the above reasons for accurate recognition.

[0030] Further, in the S4 scene recognition step, by pushing the shutter to block the flash, a grid light of alternating light and dark is formed.

[0031] Such a method has a simple structure and is convenient to implement. Description of the Drawings

[0032] Figure 1 It is a logic block diagram of the first embodiment of a real-time fall event detection method based on a surveillance video;

[0033] Figure 2 It is a structural schematic diagram of a device for emitting structured light in the third embodiment of a real-time fall event detection method based on a surveillance video. Detailed Embodiment

[0034] The following is further detailed through specific embodiments:

[0035] The marks in the drawings of the specification include: flash 100, rack 101, limit block 102, thermal cylinder 103, trigger switch 104, camera 201, drive motor 301.

[0036] Example 1

[0037] A real-time fall event detection method based on surveillance video in this example (as Figure 1 shown) includes:

[0038] S1 Image acquisition step: Real-time collect video image information of the target area, subtract the pixel value C k (x, y) of the point with coordinates (x, y) in the k-th frame of the video image information from the pixel value Bk(x, y) of the corresponding point in the background image to obtain the difference image D k (x, y), and compare the difference image D k (x, y) with the set threshold T to obtain the binary image R k (x, y);

[0039] S2 Human pose recognition step: Use the RVM classifier for the binary image R k (x, y) to determine the target pose category, and then input the target pose category as a target pose value sequence into the HMM model for evaluation to obtain the posterior probability of the change in the target pose category. The target pose categories include: standing, tilting, lying flat, and others;

[0040] S3 Image calibration step: Manually calibrate all rest areas in the video image information to obtain the rest area image X k (x, y);

[0041] S4 Scene recognition step: If the posterior probability is greater than the predetermined threshold, determine that the target is suspected of falling, then turn on the flash 100, determine whether the target head pose has changed according to the brightness change of the video image information. If the target head pose has not changed, it is judged as a fall. If the target head pose has changed, emit alternating bright and dark grid lights to the target area, collect the video image information of the target area again, and determine the depth relationship between the binary image R k (x, y) and the rest area image X k (x, y) to determine whether the target and the rest area are at the same depth. If they are at the same depth and their abscissas coincide, it is judged that the target has not fallen; if the target and the rest area are not at the same depth, it is judged that the target is a secondary suspected fall;

[0042] S5 Final determination of human pose step: If it is judged that the target is a secondary suspected fall, determine the direction in which the target head is facing according to the video image information. If there is a high-brightness area in the facing direction, it is judged that the target has not fallen; if the target head is facing the ground or the sky, it is judged that the target has fallen.

[0043] The RVM classifier includes three binary classifiers. Among them, the first binary classifier distinguishes the first three of the four categories from the fourth category, the second binary classifier distinguishes the first two of the four categories from the third category, and the third binary classifier distinguishes the second category from the first category among the four categories.

[0044] The training process of the RVM classifier includes the following steps:

[0045] Step 11: Extract the shape and position features from the training sample videos belonging to the four categories respectively to form a feature matrix X, and the target pose categories corresponding to all the training sample videos form a vector Y;

[0046] Step 12: Use the RBF kernel function to map the feature matrix X to a high-dimensional space;

[0047] Step 13: According to the Bayesian criterion, solve the optimal weight distribution and distribution parameters as the RVM parameters.

[0048] In specific use: The real-time fall detection method based on HMM and RVM mainly includes the following steps in the model training stage: 1) Feature extraction, which is used to extract multiple features reflecting the human body's pose changes from the training video frames of two cameras at different angles; 2) Pose classification, using the extracted above features to train a classifier, and obtaining the pose category of each training video frame through the classifier; 3) Using the hidden Markov model (HMM) to model the pose changes during the fall process (generate the HMM model).

[0049] The real-time fall detection method based on RVM and HMM mainly includes the following steps in the event detection stage: 1) Feature extraction, which is used to extract multiple features reflecting the human body's pose changes from the test video frames of two cameras at different angles; 2) Pose classification, obtaining the pose category of each test video frame using the trained classifier; 3) Using the HMM model generated in the above training stage to evaluate each movement process (the process of pose category change) in the surveillance video, so as to judge whether a fall event occurs.

[0050] By decomposing the fall process, the human body's pose change process can be attributed to standing - tilting - lying (on the ground). Based on this, using the features composed of the human body's geometric appearance and scene information as the input of the RVM classifier to judge the pose, the human body's pose in home videos can be roughly divided into four categories: standing, tilting, lying flat, and others.

[0051] In specific use, the posture of the human body is determined by judging the human body height ratio, the human body center position, and the effective area ratio. For example, the target of the human body is extracted by the minimum bounding rectangle. When the aspect ratio of the human body is greater than the threshold Th, it is determined that the human body is in a standing state. The threshold Th will vary according to personal factors, etc., and the value range is [1.5, 2]. In this embodiment, the value is taken as 1.8.

[0052] In S1, through a method similar to frame difference method, the difference image is obtained by comparing and subtracting the real-time video image information with the background image (pre-set or an image that has not changed for more than the preset time collected in advance, and the preset time can be a time period or a duration). Then, the difference image is further processed to obtain a binary image. The binary image obtained in this way is convenient for subsequent posture processing and recognition.

[0053] After obtaining the binary image, in fact, the contour information of the target is obtained, and it is used by the RVM classifier for target posture recognition to determine the target posture category. Then it is input into the HMM model for evaluation. In short, in fact, the state information of the target posture is determined by recognizing the aspect ratio and skeleton of the binary image. In fact, the above process already has mature technologies in the prior art and will not be elaborated here. The foregoing technology is essentially the recognition of the state of the target. Although the state of the target can be obtained, it cannot be determined whether it is caused by a fall or the achievement of the target's own will.

[0054] Therefore, on the basis of the above, this solution adds the calibration of the rest area, which is actually pre-calibrated manually, that is, the user (the guardian of the target) can determine items such as sofas, chairs, beds, etc. as the rest area by clicking. When it is determined that the target is suspected of falling, the flash 100 is activated. At this time, if the target has fallen, the target's attention should be immersed in self-examination or cause investigation. At this time, the target's reaction to external stimuli will be relatively weak, and the stimulation of the flash 100 will not attract the target's attention, and the target's head will basically remain unchanged. Therefore, when it is judged as a fall at this time, it can be quite reasonable.

[0055] Of course, using whether there is an emergency response to the flash 100 as the only criterion is still prone to problems. Therefore, even if there is an emergency response, in this embodiment, a grid light with alternating light and dark will be sent to obtain depth information for reconfirmation. By collecting the grid distortion, it can be determined whether the target is located in the rest area. If it is located in the rest area, it is judged that the target actively approaches the rest area and changes its posture to rest, so it is judged that the target has not fallen. If it is not in the rest area, the target may have fallen or there may be other reasons, so further investigation is needed.

[0056] In this embodiment, for such a situation, the head orientation of the target is obtained again. If there is a highlighted area in the facing direction of the head, it is usually a device such as a mobile phone or a TV. The target may change its posture to use the relevant device. Therefore, the risk is excluded and it is determined that the target has not fallen. If the target's head is facing the ground (most likely tripped) or facing the sky (most likely fallen or slipped, because the judgment of whether it is in the rest area has been made before, so it can be determined that the target did not lie down actively), it can be determined that the target has fallen. This reduces the situation of false positives and missed reports.

[0057] Based on the determination that the target is suspected of falling, this embodiment adds a combined judgment of the rest area, whether there is a highlight in the direction the head is facing, and the final direction the head is facing, which can accurately exclude high-frequency behaviors of the target such as watching TV, playing with mobile phones, sleeping, etc., and avoid false positives and missed identifications.

[0058] Embodiment 2

[0059] The difference between this embodiment and Embodiment 1 is that this embodiment further includes S4-1 for identifying high-risk areas, and determining the standing posture and rapidly changing areas in the binary image R k (x, y) and the manually calibrated channel area as high-risk areas. If it is determined that the suspected fall does not occur in the high-risk area, it is determined that the target has fallen. In the S4 scene recognition step, after the flash 100 is activated, if a highlighted area is captured in the binary image R k (x, y), it is determined that the target has fallen. In the S4 scene recognition step, by pushing the shutter to block the flash 100, an alternating light and dark grid light is formed.

[0060] In actual use, the alternating light and dark grid light is actually equivalent to structured light. The structured light is projected onto the scene, and the image sensor captures the corresponding pattern with the structured light. Since the pattern of the structured light will be deformed due to the shape of the object, the depth information of each point in the scene can be calculated using the triangulation principle based on the position and deformation degree of the pattern image in the captured image. This technology belongs to the prior art and will not be elaborated here.

[0061] Embodiment 3

[0062] Compared with Embodiment 2, the difference is only that the alternating light and dark grid light is realized through the following device (such as Figure 2As shown in the figure, the device includes: a flash 100, a fixing bracket, a driving motor 301, a rack 101, a thermal cylinder 103, a trigger switch 104, and a single-chip microcomputer. The single-chip microcomputer is consistent with the trigger signal of the flash 100. After the flash 100 receives the start signal, the single-chip microcomputer also receives the start signal simultaneously, and the single-chip microcomputer controls the driving motor 301 to rotate. The fixing bracket is L-shaped, with the upper end fixed to the camera 201 (hardware for realizing the image acquisition function). The driving motor 301 is fixed on the fixing bracket. There is a through hole on the fixing bracket for the longitudinal passage of the rack 101. Limit blocks 102 are fixedly arranged at both the upper and lower ends of the rack 101. The top of the rack 101 is fixedly connected to the thermal cylinder 103. There is a working medium that expands when heated in the thermal cylinder 103. In this embodiment, clear water is selected. A trigger switch 104 is arranged at the upper end of the thermal cylinder 103, and the trigger switch 104 is signal-connected to the single-chip microcomputer. A gear is arranged on the output shaft of the driving motor 301, and the gear meshes with the rack 101. The rack 101 is slidably connected to the fixing bracket, and a light-shielding cloth with alternating light and dark is fixedly arranged on the rack 101.

[0063] During specific use: When the single-chip microcomputer receives the signal triggered by the flash 100, which can actually be achieved through a photosensitive resistor, the single-chip microcomputer sends a first control signal to the driving motor 301, and the driving motor 301 rotates forward, driving the rack 101 to move upward. Since the light-shielding cloth is fixedly arranged on the rack 101, at this time, light blocks with alternating light and dark are emitted. Although the LED (flash 100) is a cold light source, the heat of its related circuit is still very high. And in this embodiment, due to the target falling, from the perspective of facilitating user use, the flash 100 should actually be kept in the working state to assist the user in lighting and facilitate obtaining the light intensity of the factual image (the reason for the target falling may even be tripping due to too dark light). In this way, the obtained real-time image information is clearer and convenient to transmit to relevant personnel.

[0064] Therefore, since the long-term use of the LED lamp will cause a large amount of heat generation, if the light-shielding cloth always blocks, it will affect the air flow and the heat dissipation effect. Therefore, after a large amount of heat accumulates, the clear water in the thermal cylinder 103 (of course, it can also be replaced with other liquids with a lower boiling point, such as alcohol, and those skilled in the art can choose according to the actual situation) expands when heated, causing the trigger switch 104 to be triggered. At this time, the single-chip microcomputer receives the signal and controls the driving motor 301 to reverse, causing the rack 101 to move downward to prevent the light-shielding cloth from affecting the heat dissipation and image acquisition effects.

[0065] The above are only embodiments of the present invention. The invention is not limited to the fields involved in this embodiment. Common knowledge such as specific structures and characteristics known in the art is not described in detail here. Those of ordinary skill in the art know all the common general technical knowledge in the technical field to which the invention pertains before the filing date or the priority date, are able to know all the prior art in this field, and have the ability to apply conventional experimental means before this date. Those of ordinary skill in the art can, under the inspiration given in this application, combine their own abilities to improve and implement this solution. Some typical well-known structures or well-known methods should not become obstacles for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can also be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application shall be subject to the content of its claims, and the specific implementation manners and the like recorded in the specification can be used to interpret the content of the claims.

Claims

1. A real-time fall event detection method based on surveillance video, comprising: S1 Image acquisition step: Real-time acquisition of video image information of the target area. Subtract the pixel value Ck(x, y) of the point with coordinates (x, y) in the k-th frame of the video image information from the pixel value Bk(x, y) of the corresponding point in the background image to obtain a difference image Dk(x, y). Compare the difference image Dk(x, y) with a set threshold T to obtain a binary image Rk(x, y); S2 Human pose recognition step: Use the RVM classifier for the binary image Rk(x, y) to determine the target pose category. Then input the target pose category as a target pose value sequence into the HMM model for evaluation to obtain the posterior probability of the change in the target pose category. The target pose categories include: standing, tilting, lying flat, and others; S3 Image calibration step: Manually calibrate all rest areas in the video image information to obtain a rest area image Xk(x, y); S4 Scene recognition step: If the posterior probability is greater than a predetermined threshold, determine that the target is suspected of falling, then turn on the flash. According to the brightness change of the video image information, determine whether the target head pose has changed. If the target head pose has not changed, it is judged as a fall. If the target head pose has changed, emit an alternating bright and dark grid light to the target area, and acquire the video image information of the target area again. Determine the depth relationship between the binary image Rk(x, y) and the rest area image Xk(x, y) according to the grid distortion situation, and determine whether the target and the rest area are at the same depth. If they are at the same depth and their abscissas coincide, it is judged that the target has not fallen; after turning on the flash, if a high-brightness area is acquired in the binary image Rk(x, y), it is determined that the target has fallen. If the target and the rest area are not at the same depth, it is judged that the target is a secondary suspected fall; S5 Final determination of human pose step: If it is judged that the target is a secondary suspected fall, determine the direction in which the target head is facing according to the video image information. If there is a high-brightness area in the facing direction, it is judged that the target has not fallen; if the target head is facing the ground or the sky, it is judged that the target has fallen; The RVM classifier includes 3 two-classifiers. Among them, the first two-classifier distinguishes standing, tilting, lying flat, and others in the target pose category, the second two-classifier distinguishes standing, tilting, and lying flat in the target pose category, and the third two-classifier distinguishes tilting and standing in the target pose category; The training process of the RVM classifier includes the following steps: Step 11: Extract shape and position features from the training sample videos belonging to the target pose categories respectively to form a feature matrix X, and the target pose categories corresponding to all the training sample videos form a vector Y; Step 12: Use the RBF kernel function to map the feature matrix X to a high-dimensional space; Step 13: According to the Bayesian criterion, solve the optimal weight distribution and distribution parameters as the RVM parameters; It also includes identifying high-risk areas in S4, and determining the areas with standing postures and rapid changes in the binary image Rk(x, y) and the manually calibrated channel areas as high-risk areas. If it is determined that no suspected fall occurs in the high-risk areas, it is determined as the target fall.

2. The real-time fall event detection method based on surveillance video according to claim 1, wherein, In the S4 scene recognition step described above, by pushing the shutter to block the flash, a grid of light and dark is formed.

Images