A passenger behavior recognition method based on support vector machine

Abstract

The application discloses a passenger behavior recognition method based on a support vector machine, which comprises the following steps: S1, detecting the state of an elevator door by an infrared device; S2, monitoring the state of human body key points by a camera; S3, analyzing and comparing the coordinate difference of the key points; S4, feeding back signals to the system according to the coordinate difference and a vector machine classification model signal processor; and S5, recognizing the passenger behavior and processing after the system receives the signals, so as to solve the problem of the detection precision of the existing infrared device and further ensure the safety of passengers in the elevator when accidents occur.

Description

Technical Field

[0001] This invention relates to the fields of infrared recognition and image processing for elevator equipment, and particularly to a passenger behavior recognition method based on support vector machines. Background Technology

[0002] Data shows that elevators, serving specific floors within a building, have become an indispensable mode of transportation in people's lives. With their increasingly frequent use, elevator safety has become a major concern. Advances in elevator infrared recognition devices and image recognition technology have effectively reduced the occurrence of elevator accidents involving people being trapped.

[0003] Infrared devices detect people or objects by creating barriers. When a person or object passes through an elevator door, the infrared detection device can effectively identify the barrier and send a signal to the control system to maintain the door open or send an open signal. However, infrared devices can only detect barriers, and in most elevators, the infrared device is integrated with the door system, making removal or addition cumbersome. Therefore, elevator doors lack additional infrared detection at low heights. Consequently, the system cannot effectively detect objects passing through the elevator door at low heights or people falling, potentially leading to accidents.

[0004] Human pose recognition is a research area in computer vision, currently applied in fields such as human motion analysis and video surveillance. Recognizing human motion greatly benefits our daily lives, for example, in pedestrian fall detection, sports activity analysis, and health risk assessment. The goal of motion recognition is to identify key body parts in a video. Since a video can be viewed as a data structure composed of frames arranged over time, we can analyze each frame to determine the specific actions of the person in the video.

[0005] Chinese patent document CN115527271A discloses a "System and Method for Detecting Abnormal Passenger Behavior in Elevator Cars". This invention constructs an elevator network system and introduces intelligent elevator video monitoring. The system uses network cameras inside the elevator car to capture real-time video information of passengers. Through intelligent image processing technologies, it performs in-depth analysis of the acquired video and trains a model using deep learning algorithms to achieve accurate and rapid identification of abnormal passenger behaviors such as violent conflicts and falls in the elevator, providing real-time and accurate safety assurance for elevator passengers. However, relying solely on video monitoring may lead to misjudgments, and the process for handling abnormal situations is lengthy and inefficient. Summary of the Invention

[0006] This invention primarily addresses the technical problems of low efficiency and low accuracy in existing methods for monitoring and processing passenger behavior in elevators. It provides a passenger behavior recognition method based on support vector machines, which identifies the behavioral state of passengers in the elevator by setting up infrared detection devices and cameras inside the elevator, and then accurately judges the passenger's condition and obtains an accurate processing method by combining the vector machine model.

[0007] The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions: The present invention includes: The S1 infrared device detects the status of the elevator door; The S2 camera monitors the status of key points on the human body; S3 analysis compares the coordinate differences of key points; S4 feeds back signals to the system based on the coordinate difference and the vector machine classification model signal processor; After receiving the signal, the S5 system identifies the passenger's behavior and takes appropriate action.

[0008] An elevator recognition device for identifying passenger actions includes a detachable infrared transmitter and receiver, a signal converter, and a method combining human body key point recognition. The infrared device can determine the state between the passenger and the elevator door; based on infrared beam obstruction, it can determine whether the passenger is blocking the elevator door. Then, the coordinates of human body key points and a vector machine classification model are used to determine the passenger's behavioral state. Finally, based on the state analysis, a signal processor sends corresponding feedback signals to the system, prompting the system to make appropriate decisions based on the actual situation.

[0009] Preferably, in step S1, the infrared device can detect two situations: one is that only the lower half of the infrared light emitted by the infrared transmitter installed on the elevator door is blocked, while the upper half can be received normally; the other is that the infrared light received by the infrared receiver changes to only the upper half being receivable. In either of these situations, the signal converter connected to the infrared transmitter and receiver converts the signal fed back by the infrared device to one; otherwise, the fed-back signal is zero. If a person is abnormally standing inside the elevator, the infrared light emitted by the infrared device near the bottom of the elevator door will be blocked by the person. Therefore, the transmission and reception status of the infrared device near the bottom (the lower half) can be used to preliminarily determine whether the passenger's behavior is abnormal.

[0010] Preferably, in step S2, a camera installed in the elevator identifies the head, hands, hips, and ankles as key points for recognition. These key points are placed in the same three-dimensional coordinate system, with each key point having three coordinates: X, Y, and Z. The positive x-direction corresponds to the direction of exiting the elevator. The camera captures images of the actions of passengers entering the elevator. Based on the pre-set coordinate values ​​of the key points, the current state of the passenger is determined by comparing these coordinate values. When a condition is met, the signal converter returns a one signal; otherwise, it returns zero. By setting the main body parts as key points and establishing them in the same coordinate system, the differences and changes between different coordinate values ​​in the coordinate system can more intuitively distinguish changes in passenger behavior, reducing misjudgments of passenger behavior.

[0011] Preferably, in step S2, the head coordinates are set as head_X, head_Y, and head_Z; the hand coordinates are set as hand_X, hand_Y, and hand_Z; the hip coordinates are set as hip_X, hip_Y, and hip_Z; and the ankle coordinates are set as ankle_X, ankle_Y, and ankle_Z. When the Z-coordinates of the four key points are all close to zero, and head_X and hand_X are less than zero, while hip_X and ankle_X are greater than zero, the signal processor will feed back a signal value of 1. When the system receives two consecutive 1 signal values, the system determines that an accident has occurred and takes event handling measures. The system may sometimes misjudge changes in coordinate values. If the system only receives one 1 signal value from the signal processor and judges that the passenger behavior is abnormal, causing the elevator to stop, it is very likely to make a misjudgment and affect the normal operation of the elevator. Receiving two consecutive 1 signal values ​​from the signal processor to confirm the passenger behavior is abnormal can further improve the accuracy of the system's judgment of passenger behavior.

[0012] Preferably, in step S2, a support vector machine (SVM) classification model is used to identify and judge the specific gate-blocking posture. Simultaneously, based on the SVM classification model, a "one-to-many" method is used to classify possible events into four categories, and a binary SVM is constructed to distinguish the i-th category from the other categories. Where the subscript t represents the sample index; the superscript i represents the category to be identified; a total of 4 SVMs were trained. The decision function for the i-th class is:

[0013] w is the normal vector, which determines the direction of the hyperplane to be identified. The superscript i indicates the type of recognition, the subscript t is the index of the sample, Φ represents the nonlinear mapping from the input space to the feature space, and X... t Let be the t-th sample. Where, the second term... The part introduced for soft-margin support vector machines allows some sample points to not satisfy the condition of margin greater than or equal to 1 when recognizing human action position information. The penalty coefficient C>0, and the slack variable ξ≥0. A total of 4 SVMs are trained. Using the Lagrange multiplier method, the classification decision function for the i-th class is derived. In the decision function, where d... i Let be the parameters of the i-th class of binary SVM, α be the Lagrange multiplier vector, and y be the parameter of the binary SVM. t For sample point X t The class label. This prediction function can be viewed as the new data X. new The signed functional interval to the i-th decision boundary, since the sample points that meet the condition are set to class +1, when X new When the i-th class is selected, the classification decision function is positive.

[0014] Preferably, in step S2, for the first category, b1, where a passenger's leg obstructs the normal opening of the elevator door, the judgment criteria are: head_X, hip_X, and hand_X are all greater than zero, while ankle_X1 or ankle_X2 is less than zero. When a leg obstructs the normal opening of the elevator door, the main change is in the coordinate value of the ankle, so judging the ankle value helps the system confirm whether this situation has occurred. Conventional two-class classifiers can only answer positive and negative questions, making + and - judgments on the results. However, real-world problems are often not limited to just positive and negative, so a two-class classifier can better encompass most situations, enabling the system to make accurate judgments and take corresponding actions for more situations.

[0015] Preferably, in step S2, for the second type of situation (b2) where a passenger's body obstructs the normal opening of the elevator door, the judgment criteria are: head_X, hip_X, hand_X, and ankle_X are all close to and near zero, with a danger range of [-10, 10]. When a body obstructs the normal opening of the elevator door, the coordinate values ​​of the four key points will change. Therefore, judging the coordinate values ​​of all key points can help the system confirm whether this situation has occurred.

[0016] Preferably, in step S2, for the third type of situation (b3) where a passenger falls and obstructs the normal opening of the elevator door, the judgment criteria are: head_X and hand_X are less than zero; hip_X and ankle_X are greater than zero. When the head obstructs the normal opening of the elevator door, the main changes are in the coordinate values ​​of head and hand. Therefore, judging the head and hand values ​​can help the system confirm whether this situation has occurred.

[0017] Preferably, in step S2, for the fourth type of situation (b4) where a passenger's arm obstructs the normal opening of the elevator door, the judgment criterion is: the difference between hand_Z and head_Z changes, and the value of hand_X is less than the other three types. When an arm obstructs the normal opening of the elevator door, the main change is in the coordinate value of hand; therefore, judging the hand value can help the system confirm whether this situation has occurred.

[0018] Preferably, in step S1, an infrared transmitter and an infrared receiver are installed at the ground-facing end of the elevator door and the corresponding ground-facing end of the car wall. Since abnormal passenger behavior obstructing the opening and closing of the elevator door primarily occurs at the ground-facing end of the elevator door, installing infrared transmitters and receivers at the ground-facing ends of the elevator door and the corresponding ground-facing ends of the car wall improves the accuracy of the infrared devices in detecting passenger behavior.

[0019] The beneficial effects of this invention are: by installing an infrared device in an existing elevator and combining it with human key point recognition technology to recognize when a person falls and other actions, the accuracy of judging human-related behaviors can be improved and the time for judging behaviors can be shortened. Attached Figure Description

[0020] Figure 1 This is a flowchart of a passenger behavior recognition method based on support vector machines according to the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the classification calculation using the vector machine classification model.

[0022] Figure 3 This is a schematic diagram of the vector machine classification algorithm. Detailed Implementation

[0023] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0024] Example: This example illustrates a passenger behavior recognition method based on support vector machines, such as... Figure 1 , Figure 2 and Figure 3As shown, the process includes: S1, an infrared device detecting the elevator door status; S2, a camera monitoring the status of key human body points; S3, analyzing and comparing the coordinate differences of the key points; S4, a signal processor feeding back a signal to the system based on the coordinate differences and the vector machine classification model; and S5, the system receiving the signal, recognizing passenger behavior, and processing it. In step S1, the infrared device can detect two situations: one is that only the lower half of the infrared light emitted by the infrared transmitter installed on the elevator door is blocked, while the upper half can be received normally; the other is that the infrared light received by the infrared receiver changes so that only the upper half can be received. In both cases, the signal converter connected to the infrared transmitter and receiver converts the signal fed back by the infrared device to one; otherwise, the fed-back signal is zero. In step S2, the head, hands, hips, and ankles of the human body are identified by a camera installed in the elevator. These human body key points are set in the same three-dimensional coordinate system, and each key point has three coordinates: X, Y, and Z. The positive x-direction is the direction of exiting the elevator. The camera acquires the motion image of the passenger entering the elevator. Based on the acquired coordinate values ​​of the pre-set human body key points, the current state of the passenger is determined by comparing the coordinate values. When the condition is met, the signal converter returns a signal of one; otherwise, it returns a signal of zero. In step S2, the head coordinates are set as head_X, head_Y, and head_Z; the hand coordinates are set as hand_X, hand_Y, and hand_Z; the hip coordinates are set as hip_X, hip_Y, and hip_Z; and the ankle coordinates are set as ankle_X, ankle_Y, and ankle_Z. When the Z-coordinate values ​​of the four key points are close to zero, and head_X and hand_X are less than zero, while hip_X and ankle_X are greater than zero, the signal processor will feed back a signal value of 1. When the system receives two consecutive 1 signal values, the system determines that an accident has occurred and takes event processing measures. In step S2, a support vector machine (SVM) classification model is used to identify and judge the specific door-blocking posture. Based on the SVM classification model, the possible events are divided into four categories using a "one-to-the-others" method. A binary SVM is constructed to distinguish the i-th category from the other categories. Where the subscript t represents the sample index; the superscript i represents the category to be identified; a total of 4 SVMs were trained. The decision function for the i-th class is: In step S2, for the first category, b1, where a passenger's leg obstructs normal door opening, the judgment criteria are: head_X, hip_X, and hand_X are all greater than zero, while ankle_X1 or ankle_X2 is less than zero. In step S2, for the second category, b2, where a passenger's body obstructs normal door opening, the judgment criteria are: head_X, hip_X, hand_X, and ankle_X are all close to and near zero, with a danger range of [-10, 10]. In step S2, for the third category, b3, where a passenger falls and obstructs normal door opening, the judgment criteria are: head_X and hand_X are less than zero; hip_X and ankle_X are greater than zero. In step S2, for the fourth category, b4, where a passenger's arm obstructs normal door opening, the judgment criteria are: the difference between hand_Z and head_Z changes, and the value of hand_X is less than the other three. In step S1, infrared transmitters and receivers are installed at the ground-facing ends of the elevator doors and the corresponding ground-facing ends of the car walls. Infrared radiation is emitted by an infrared transmitter and received by an infrared receiver. This receiver is easily detachable and reassembled. The device primarily detects two scenarios: 1. When obstruction changes from complete obstruction to partial obstruction (lower half only); 2. When obstruction occurs only in the lower half. In either of these scenarios, the signal converter converts the detected infrared radiation into a signal of "one"; otherwise, it returns a signal of "zero." Combined with methods for recognizing key human body points, when... Figure 3In the scenario shown, the coordinates of the four corresponding recognition points satisfy the following conditions: head_X and hand_X are less than zero, hip_X and ankle_X are greater than zero, and the Z-values ​​of the ordinates of head_Z, hip_Z, hand_Z, hip_Z, and ankle_Z are approximately zero. If these conditions are met, an event is determined to have occurred, and a signal is fed back to the signal processor; otherwise, the result is zero. When the signal converter receives two one-signals, it determines that a fall has occurred and sends a delayed door-opening signal to the system. A support vector machine (SVM) classification model is used to identify and judge specific door-blocking postures, achieving recognition of multiple categories of passenger actions. The model is trained on a sample set of judged recognition cases to achieve passenger action recognition. A binary classifier only answers positive and negative questions, making + and - judgments on the results. However, real-world problems often require multiple results for judgment, such as text recognition, photo recognition, and passenger action recognition. Therefore, this paper proposes considering multiple scenarios of passenger door blocking and solving a multi-objective function problem. After monitoring captures the passenger's entry into the elevator, the aforementioned human keypoint recognition is used to train the passenger's action model. The support vector machine (SVM) is configured to recognize four scenarios: passenger falling, passenger blocking the door with their body, passenger blocking the door with their hand, and passenger blocking the door with their foot. When making judgments, the SVM uses a "one-to-many" approach, meaning it still considers two scenarios each time. For each scenario, it evaluates: if the scenario meets the criteria, it's considered "1+"; if not, it's considered "1-". The specific judgment criteria are: for... Figure 3 For class 1 in the graph, all others are class -1. Construct the corresponding binary SVM; its decision boundary is b1 in the graph. Figure 3 If there are 2 classes, then the rest are -1 classes. Construct the corresponding binary SVM with a decision boundary of b2; for Figure 3 For the three classes in the diagram, all others are of class -1. Construct the corresponding binary SVM; its decision boundary is b3 in the diagram. Figure 3 For the four classes in the diagram, all others are classified as -1. A corresponding binary SVM is constructed, and its decision boundary is b4 in the diagram. Different binary SVMs are constructed for different classes, and a quadratic programming problem is used to separate the i-th class from the others. This patent trains a total of four SVMs to address the existing problem. Each class uses a corresponding decision function. Through decision conditions, combined with the previous support vector machine algorithm, the system can recognize the actions of passengers entering the elevator, ensuring that passengers can safely enter the elevator and taking timely measures in case of an accident.

Claims

1. A passenger behavior recognition method based on support vector machines, comprising the following steps: The S1 infrared device detects the elevator door status and is configured to detect two states: one is that only the lower half of the infrared light emitted by the infrared transmitter installed on the elevator door is blocked, while the upper half of the infrared light can be received normally; the other is that the infrared light received by the infrared receiver changes to only the upper half being receivable. When either of these two states occurs, the signal converter connected to the infrared transmitter and receiver converts the signal fed back by the infrared device into one; otherwise, the fed back signal is zero. The S2 camera monitors the status of key points on the human body, including the head, hands, hips, and ankles, which are set in the same three-dimensional coordinate system, where the positive X direction is the direction of exiting the elevator; S3 analyzes and compares the coordinate differences of key points to determine the current status of passengers. S4 feeds back signals to the elevator control system based on the coordinate difference and the vector machine classification model signal processor. The vector machine classification model uses a pair of other methods to classify events. After receiving the signal, the S5 elevator control system identifies the passenger's behavior and takes appropriate action. In S2, the coordinates of the head are set as head_X, head_Y, and head_Z; the coordinates of the hands are set as hand_X, hand_Y, and hand_Z; the coordinates of the hips are set as hip_X, hip_Y, and hip_Z; and the coordinates of the ankles are set as ankle_X, ankle_Y, and ankle_Z. When the Z-coordinates of the four key points are close to zero, and head_X and hand_X are less than zero, while hip_X and ankle_X are greater than zero, the signal processor feeds back a signal value of one. When the elevator control system receives two consecutive one signal values, the elevator control system determines that an accident has occurred and takes event handling measures.

2. The passenger behavior recognition method based on support vector machine according to claim 1, characterized in that, In step S2, a support vector machine (SVM) classification model is used to identify and judge specific gate-blocking postures. Simultaneously, based on the SVM classification model, a "one-to-many" method is used to classify possible events into four categories. A binary SVM is constructed to distinguish the i-th category from the other categories. ， ， ， ， Where the subscript t is the sample index; the superscript i is the category to be identified, and the decision function for the i-th category is: ， In the formula, w is the normal vector, the superscript i represents the category being recognized, the subscript t represents the sample index, and Φ represents the nonlinear mapping from the input space to the feature space. For the t-th sample, the second term This is a part introduced for soft-margin support vector machines, with a penalty coefficient C > 0, and slack variables. A total of 4 SVMs were trained, and in the decision function, Let be the parameters of the i-th class of binary SVM, and α be the Lagrange multiplier vector. For sample points The class label, this prediction function can be viewed as new data. The signed functional interval to the i-th decision boundary, since the sample points that meet the condition are set to class +1, therefore when When the i-th class is selected, the classification decision function is positive.

3. The passenger behavior recognition method based on support vector machine according to claim 2, characterized in that, In step S2, for the first type of situation where a passenger's leg blocks the normal door opening (b1), the judgment criteria are: head_X, hip_X, and hand_X are all greater than zero, while ankle_X1 or ankle_X2 is less than zero.

4. The passenger behavior recognition method based on support vector machine according to claim 2, characterized in that, In step S2, for the second type of situation where the passenger's body obstructs the normal opening of the door (b2), the judgment criteria are: head_X, hip_X, hand_X, and ankle_X are all close to and near zero, and the danger range is [-10, 10].

5. The passenger behavior recognition method based on support vector machine according to claim 2, characterized in that, In step S2, for the third type of situation b3, where a passenger falls and obstructs the normal opening of the door, the judgment criteria are: head_X and hand_X are less than zero; hip_X and ankle_X are greater than zero.

6. The passenger behavior recognition method based on support vector machine according to claim 2, characterized in that, In step S2, for the fourth category, b4, which is a passenger's arm blocking the normal door opening, the judgment criteria are: the difference between hand_Z and head_Z changes, and the value of hand_X is less than the X value of the head coordinate, hip coordinate, and ankle coordinate.

7. The passenger behavior recognition method based on support vector machine according to claim 1, characterized in that, In step S1, an infrared transmitter and an infrared receiver are provided on the elevator door near the ground and on the corresponding car wall near the ground.

Images