A multi-target mouse behavior online recognition and body temperature monitoring method

Abstract

The present invention discloses a method for online identification of multi-target mouse behaviors and monitoring body temperature, which collects data of multi-target mouse image information and temperature information based on an infrared thermal imaging camera; pre-processes the mouse images collected by the infrared thermal imaging camera; analyzes the pre-processed mouse images based on a deep learning image processing method to obtain the body contour and key point information of each mouse; obtains the identity ID information of each mouse based on the mouse image information and body contour information based on a multi-target tracking algorithm; calculates the mouse surface temperature based on the mouse identity ID information, body contour information and the temperature information; calculates the individual behavior and social behavior of the mouse based on the key point information; and displays the body temperature information and behavior of each mouse in real time. The present invention realizes the tracking of multi-target mice and the real-time analysis of individual behavior and social behavior, and provides technical support for the analysis of multi-target mouse behavior.

Description

Technical Field

[0001] The invention belongs to the field of computer vision detection technology and mouse behavior analysis, and specifically relates to a multi-target mouse behavior online recognition and body temperature monitoring method. Background Art

[0002] Animal behavior is a discipline that studies how animals achieve a dynamic balance between their environment and behavior. Animal behavior contains a lot of health information, directly or indirectly reflecting its internal state of physiology, pathology, and nutrition. Therefore, observing and effectively identifying animal behavior is of great significance to neuroscience and drug treatment. Mice are mammals, and their genes are very similar to humans. They are easy to raise and maintain, and are one of the most commonly used animals in animal behavior experiments.

[0003] Currently, most video behavior analysis of mice is done offline, requiring the use of a server to post-process and analyze the saved videos. Real-time online analysis is not possible, and behavioral analysis can often only be performed on a single mouse. The scene is monotonous, and only some simple individual behaviors of mice can be obtained. For multiple mice, due to the inability to accurately track them, the social behavior between mice cannot be accurately analyzed.

[0004] When studying the social behavior of mice, it is necessary to track multiple mice at the same time. However, it is difficult to accurately identify and track multiple mice with similar appearances, especially when there is mutual occlusion between mice, which makes it easy to track incorrectly. Traditional tracking methods require artificial marking of the mouse's body, distinguishing different mice by artificial markers, or dyeing the mouse to identify different mice based on texture information and color information, and using wireless radio frequency identification technology to obtain the mouse's position for tracking. However, marking the mouse may affect the mouse's physiological activities and thus affect the mouse's behavior, and wireless radio frequency identification technology has a complex structure and high cost. Summary of the invention

[0005] In view of the problems that the above-mentioned existing mouse behavior recognition methods can only target a single mouse, cannot obtain the social behavior between mice, cannot perform real-time online analysis and require a server for post-processing analysis, and have low efficiency, the present invention proposes a multi-target mouse behavior online recognition and body temperature monitoring method.

[0006] In order to solve the technical problem, the present invention provides a multi-target mouse behavior online recognition and body temperature monitoring method.

[0007] A method for online recognition of multi-target mouse behaviors and monitoring body temperature, characterized in that it comprises the following steps:

[0008] Step S1, collecting image information and temperature information of multiple target mice based on an infrared thermal imaging camera; transmitting the image information and temperature information to an edge computing unit;

[0009] Step S2: the edge computing unit pre-processes the mouse images collected by the infrared thermal imaging camera; and uses a deep learning-based image processing method to analyze the pre-processed mouse images to obtain the body contour and key point information of each mouse;

[0010] Step S3, obtaining the identity ID information of each mouse according to the mouse image information and body contour information based on a multi-target tracking algorithm;

[0011] Step S4, calculating the body surface temperature of each mouse according to the mouse's ID information, body contour information and the temperature information;

[0012] Step S5, calculating the individual behavior and social behavior of each mouse according to the mouse's identity ID information and key point information;

[0013] Step S6, displaying the body temperature information and behavior of each mouse on the display screen in real time.

[0014] Preferably, in the step S2, the mouse image captured by the infrared thermal imaging camera is converted from a multi-channel color image to a single-channel grayscale image, and irrelevant background areas are cropped; the body contour of each mouse is segmented using the instance segmentation network YOLOv8-seg to obtain the body contour area information of each mouse; the body key points of each mouse are detected using the YOLOv8-pose key point detection network, and the key points include the tip of the nose, left ear, right ear, neck, body center and tail root.

[0015] Preferably, in step S3, the Strongsort multi-target tracking algorithm is specifically used to initialize a track of the detected new target and assign an identity ID number. In the subsequent frames, the detected target and the existing track are matched according to the similarity of appearance features and position features, so as to distinguish the same mouse between the previous and next frames of the video;

[0016] The similarity of the appearance features is represented by the minimum cosine distance, and the specific process is as follows:

[0017] (1) Use the convolutional neural network OsNet to extract the appearance features of the detection target. The extraction range is the circumscribed rectangular box of the detection target, and a 512-dimensional feature vector is obtained;

[0018] (2) Calculate the minimum cosine distance between the feature vector of each detected target and the feature vector of each existing trajectory to form a minimum cosine distance matrix. The smaller the minimum cosine distance value, the greater the similarity of the appearance features of the two, and the more likely they are the same target.

[0019] The position feature similarity is represented by Mahalanobis distance, and the specific process is as follows:

[0020] (1) Use the Kalman filter algorithm to predict the state information and corresponding covariance matrix of the next frame based on the state information and covariance matrix of the previous frame of the existing trajectory;

[0021] (2) Based on the state information of the existing trajectory and the corresponding covariance matrix as well as the position information of the detected target, the Mahalanobis distance between the existing trajectory and the detected target is calculated to form a Mahalanobis distance matrix. The smaller the Mahalanobis distance, the smaller the position difference between the two, and the more likely they are the same target.

[0022] The minimum cosine distance matrix and the Mahalanobis distance matrix are added together with a weight of 0.95:0.05 to form a cost matrix; the smaller the value in the cost matrix, that is, the lower the cost, the more likely it is that the corresponding existing trajectory and the detection target will be matched successfully; the cost matrix is ​​input into the Hungarian algorithm to match the existing trajectory and the detection target to achieve the recognition of the same mouse before and after the video.

[0023] Preferably, the state information of the existing trajectory is in the form of an 8-dimensional vector Indicates that the location information of the detection target is ,in Indicates the coordinates of the center point of the mouse target box, is the aspect ratio of the frame, is the height of the frame, represent The velocity in image coordinates.

[0024] Preferably, the center point coordinates of the mouse target frame are set to the geometric centroid of the mouse body segmentation contour area; the aspect ratio of the frame Set to 1; the height of the box The body width is adapted to the body width of the mouse. The body width is obtained by fitting the body segmentation contour of the mouse to an ellipse using an ellipse fitting algorithm based on the least squares method. The length of the minor axis of the fitted ellipse is express, and The calculation relationship is .

[0025] Preferably, in step S4, the temperature information is a two-dimensional matrix of the same size as the image, representing the temperature value of each pixel in the image; the average temperature value of the pixel points within the mouse body contour is calculated from the mouse body contour information as the mouse surface temperature; the surface temperature of each mouse can be obtained from the mouse identity ID information; according to the change in surface temperature, when the mouse body temperature is abnormal, an alarm message is issued.

[0026] Preferably, in step S5, the movement characteristics and body shape characteristics of the mouse and the distance between the nose tip of the mouse and the diet trough and the water port are calculated according to the key point information of the mouse's body;

[0027] The motion feature is the displacement of the key points of the mouse body between the previous and next frames of the video;

[0028] The body shape features are the distances between key points on the mouse body and the angles between the lines connecting the key points, specifically:

[0029] Body length, from the tip of the nose to the base of the tail

[0030] Distance from nose tip to left ear

[0031] Distance from nose tip to right ear

[0032] Distance from left ear to neck

[0033] Distance from right ear to neck

[0034] Distance from neck to body center

[0035] Distance from the center of the body to the base of the tail

[0036] Head rotation angle, that is, the angle between the line from the center of the body to the neck and the line from the neck to the tip of the nose

[0037] Nose tip angle, that is, the angle between the line between the left ear and the nose tip and the line between the right ear and the nose tip

[0038] Body rotation angle, that is, the angle between the line from the base of the tail to the center of the body and the line from the center of the body to the neck

[0039] According to the above movement characteristics, body characteristics and the distance between the mouse nose tip and the diet trough and water drinking port, the support vector machine SVM was used to discriminate the individual behavior of the mice, and the kernel function of the support vector machine was selected as the Gaussian radial basis function; the identifiable individual behaviors included: eating, drinking, standing, sniffing, grooming, standing still and walking; the distances between the key points of the bodies of multiple mice and the head orientations were calculated to discriminate the social behaviors between the mice, and the social behaviors included: face sniffing, body sniffing, tail sniffing, chasing, fighting, and approaching; the number of occurrences and duration of the behavior of each mouse were recorded.

[0040] Compared with the prior art, the present invention has significantly positive technical effects, and its beneficial effects are at least reflected in the following aspects.

[0041] (1) The present invention uses an edge computing unit to analyze mouse images and mouse behaviors, and can analyze the collected videos and monitor the body temperature information of mice in real time. There is no need to save the videos and then use a dedicated server device to analyze them like the existing technology. When the mouse body is abnormal, such as when the body temperature drops significantly when the mouse is on the verge of death, an alarm message can be issued immediately.

[0042] (2) The present invention uses an infrared thermal imaging camera to collect images, and the images are generated based on temperature information. Therefore, the images will not change with changes in the background. The foreground mouse and the background are clearly distinguished. The interference caused by the background color similar to the mouse color caused by the use of RGB cameras in the prior art will not occur. Therefore, the difficulty of image analysis is reduced and the accuracy of mouse body recognition is improved. In addition, since the images are generated based on temperature information, the images will not change with the color of the mouse's hair. The analysis of mice of different colors can share a body contour segmentation model and a key point detection model, and there is no need to train the mouse body contour segmentation network and the key point detection network separately.

[0043] (3) The present invention can track multiple mice and analyze the individual behavior of each mouse and the social behavior between mice. In tracking multiple mice, the state information of the existing trajectory and the position information of the detection target are set as the geometric centroid of the mouse body segmentation contour area, the aspect ratio is set to 1, and the height is highly adaptive to the mouse body width. Compared with the prior art that directly uses the mouse body external frame as the mouse position information, the present method makes full use of the mouse body contour information, avoids the problem that the mouse external frames contain each other when the mice are close, and the center point of the external frame deviates from the center of the mouse body and cannot accurately represent the real position of the mouse body, thereby improving the tracking accuracy of the mouse in the case of approach and occlusion, thereby better analyzing the social behavior between mice. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of the method described in this application.

[0045] Figure 2 Schematic diagram of mouse images captured by an infrared thermal imaging camera after preprocessing.

[0046] Figure 3 Schematic diagram of mouse body segmentation and key point detection.

[0047] Figure 4 Schematic diagram of multi-target mouse tracking.

[0048] Figure 5 Schematic diagram of ellipse fitting for mouse body segmentation contour DETAILED DESCRIPTION

[0049] In order to solve the technical problem, the present invention provides a method for online recognition of multi-target mouse behaviors and monitoring body temperature. The technical solution of the present invention is further described below through specific implementation methods.

[0050] When the multi-target mouse behavior online recognition and body temperature monitoring method is implemented through hardware, the device includes a touch display screen, an infrared thermal imaging camera, a mouse breeding box, an edge computing unit, and a machine box.

[0051] The infrared thermal imaging camera is specifically Hikvision's online temperature measurement thermal imager (model: DS-2TD2037T-4 / X), which is fixed in the middle of the lid of the mouse breeding box to simultaneously obtain image information and temperature information of multiple mice from a top-down perspective. The obtained video frame rate is 25 frames per second, the image size is 384 pixels × 288 pixels, and the temperature information is a two-dimensional matrix of the same size as the image, recording the temperature value of each pixel.

[0052] The machine box is located directly above the mouse breeding box. It is used to place and protect the edge computing unit and fix the touch display screen.

[0053] The edge computing unit is NVIDIA's Jetson Orin Nano artificial intelligence development board, which has 6GB of video memory and can support the deployment of neural network models. The edge computing unit is placed in the machine box.

[0054] The touch screen is a CreateBlock LCD touch capacitive screen, embedded in the side of the box. The touch screen is used to input commands to the edge computing unit and display the results of mouse behavior analysis.

[0055] like Figure 1 As shown, it is a flow chart of the method of the present invention, comprising the following steps:

[0056] S1: Take the case of three mice in a mouse feeding box as an example. Use the touch screen to input instructions to the edge computing unit to control the infrared thermal imaging camera to collect data on the image information and temperature information of multiple target mice; and transmit the image information and temperature information to the edge computing unit.

[0057] S2: The edge computing unit preprocesses the mouse image captured by the infrared thermal imaging camera. Specifically, the mouse image captured by the infrared thermal imaging camera is converted from a multi-channel color image to a single-channel grayscale image, and irrelevant background areas are cropped. The preprocessed mouse image is as follows: Figure 2 shown.

[0058] The instance segmentation network YOLOv8-seg is used to obtain the body contour of each mouse from the preprocessed mouse images, and the key point detection network YOLOv8-pose is used to obtain the body key point information of each mouse, including the tip of the nose, left ear, right ear, neck, body center and tail root. Figure 3 yes Figure 2 The images obtained after YOLOv8-seg and YOLOv8-pose show the body contour information and key point position information of each mouse. YOLOv8-seg and YOLOv8-pose described in the present invention use n models with fewer parameters in order to speed up the image analysis.

[0059] S3: Input the mouse image information and body contour information into the multi-target tracking algorithm Strongsort. Strongsort initializes a track for the detected new target and assigns an ID number. In the following frames, the detected target and the existing track are matched according to the similarity of appearance features and position features, and the same mouse is distinguished between the previous and next frames of the video. Figure 4 As shown in the figure, the image on the left is the front frame of the video, and the image on the right is the back frame of the video. The same mouse in the front and back frames is assigned the same identity ID, thereby achieving the effect of multi-target mouse tracking.

[0060] The similarity of the appearance features is characterized by the minimum cosine distance, and the specific process is as follows:

[0061] (1) Use the convolutional neural network OsNet to extract the appearance features of the detection target. The extraction range is the circumscribed rectangular box of the detection target, and a 512-dimensional feature vector is obtained;

[0062] (2) Calculate the minimum cosine distance between the feature vector of each detected target and the feature vector of each existing trajectory to form a minimum cosine distance matrix. The smaller the minimum cosine distance value, the greater the similarity of the appearance features of the two, and the more likely they are the same target.

[0063] The position feature similarity is characterized by Mahalanobis distance, and the specific process is as follows:

[0064] (1) Use the Kalman filter algorithm to predict the state information and corresponding covariance matrix of the next frame based on the state information and covariance matrix of the previous frame of the existing trajectory;

[0065] (2) Based on the state information of the existing trajectory and the corresponding covariance matrix as well as the position information of the detected target, the Mahalanobis distance between the existing trajectory and the detected target is calculated to form a Mahalanobis distance matrix. The smaller the Mahalanobis distance, the smaller the position difference between the two, and the more likely they are the same target.

[0066] The minimum cosine distance matrix and the Mahalanobis distance matrix are added together with a weight of 0.95:0.05 to form a cost matrix; the smaller the value in the cost matrix, that is, the lower the cost, the more likely it is that the corresponding existing trajectory and the detection target will be matched successfully; the cost matrix is ​​input into the Hungarian algorithm to match the existing trajectory and the detection target to achieve the recognition of the same mouse before and after the video.

[0067] The state information of the existing trajectory is an 8-dimensional vector Indicates that the location information of the detection target is ,in Indicates the coordinates of the center point of the mouse target box, is the aspect ratio of the frame, is the height of the frame, represent The velocity in image coordinates.

[0068] The coordinates of the center point of the mouse target frame are set to the geometric centroid of the mouse body segmentation contour area; the aspect ratio of the frame Set to 1; the height of the box The body width is adapted to the body width of the mouse. The body width is obtained by fitting the body segmentation contour of the mouse to an ellipse using an ellipse fitting algorithm based on the least squares method. The length of the minor axis of the fitted ellipse is express, and The calculation relationship is .

[0069] Figure 5 This is the ellipse and its minor axis after elliptical fitting of the mouse body contour. It can be seen from the figure that the minor axis of the fitted ellipse is basically close to the width of the mouse body.

[0070] S4: Calculate the sum of the temperature values ​​corresponding to each pixel point in the body contour of the mouse based on the body contour information and temperature information of the mouse, and calculate the average value as the body surface temperature of the mouse. The body surface temperature of each mouse can be obtained based on the mouse ID information. The body surface temperature of each mouse is recorded. When the body surface temperature of a mouse changes abnormally, an alarm message is sent to the touch screen.

[0071] S5: Calculate the movement characteristics, body shape characteristics, and distances between the nose tip and the food trough and water port of the mouse based on the key points of the mouse's body;

[0072] The motion features are the displacement of the key points of the mouse body between the previous and next frames of the video, including: the displacement of the nose tip compared to the previous frame, the displacement of the left ear compared to the previous frame, the displacement of the right ear compared to the previous frame, the displacement of the neck compared to the previous frame, the displacement of the center of the body compared to the previous frame, and the displacement of the base of the tail compared to the previous frame. Motion features are the key to distinguishing between walking and stationary behaviors of mice.

[0073] Posture characteristics are the distances between key points of the mouse body and the angles between the lines connecting the key points, including body length, i.e., the distance from the nose tip to the base of the tail, the distance from the nose tip to the left ear, the distance from the nose tip to the right ear, the distance from the left ear to the neck, the distance from the right ear to the neck, the distance from the neck to the center of the body, the distance from the center of the body to the base of the tail, the head rotation angle, i.e., the angle between the line from the center of the body to the neck and the line from the neck to the nose tip, the nose tip angle, i.e., the angle between the line from the left ear to the nose tip and the line from the right ear to the nose tip, and the body rotation angle, i.e., the angle between the line from the base of the tail to the center of the body and the line from the center of the body to the neck. Posture characteristics reflect the body posture of the mouse and are the key to distinguishing the behavior of the mouse.

[0074] Based on the movement characteristics, body characteristics, and the distance between the mouse nose tip and the food trough and water drinking port, the machine learning algorithm support vector machine (SVM) was used to discriminate the individual behavior of the mice. The kernel function was selected as the Gaussian radial basis function, and the SVM was set to 7 categories, representing the 7 individual behaviors of mice: eating, drinking, standing, sniffing, grooming, standing still, and walking.

[0075] Calculate the distances between key points of the bodies of multiple mice and the orientation of their heads to identify the social behaviors between mice, including face sniffing, body sniffing, tail sniffing, chasing, fighting, and approaching.

[0076] The specific method for judging face sniffing, body sniffing and tail sniffing behaviors is as follows: first, determine whether there is any overlapping part between the body contours of the two mice, and then determine the direction of the mouse heads based on the key points of the mouse nose tips and necks. If the body contours of the two mice overlap and one mouse's head is facing the other, it means that one mouse is sniffing the other mouse. Then, determine the sniffing position based on the distance between the nose tip of the actively facing mouse and the key points of the body of the faced mouse. If the nose tip of the actively facing mouse is closest to the nose tip of the faced mouse, it is face sniffing; if the nose tip of the actively facing mouse is closest to the body center of the faced mouse, it is body sniffing; if the nose tip of the actively facing mouse is closest to the base of the tail of the faced mouse, it is tail sniffing.

[0077] The specific method for judging chasing behavior is as follows: if the individual behaviors of the two mice are walking, and the face of one mouse is facing the tail of the other mouse, and the distance between the nose tip of the actively facing mouse and the base of the tail of the facing mouse is less than 10 pixels, it is a chasing behavior.

[0078] The specific method for judging fighting behavior is as follows: the body contours of the two mice have overlapping areas, the two mice are in motion and the speed is greater than 10 cm / s, and their heads are facing each other's body parts.

[0079] The number of occurrences and corresponding durations of all individual behaviors and social behaviors of mice were saved to facilitate further statistical analysis of mouse behavior.

[0080] S6: Display the mouse behavior and body temperature information in real time on the touch screen.

[0081] The specific implementation cases described in this application are merely examples of the main ideas of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. A multi-target mouse behavior online recognition and body temperature monitoring method, characterized in that: The following steps are involved: Step S1, collecting image information and temperature information of multiple target mice based on an infrared thermal imaging camera; transmitting the image information and temperature information to an edge computing unit; Step S2: the edge computing unit pre-processes the mouse images collected by the infrared thermal imaging camera; and uses a deep learning-based image processing method to analyze the pre-processed mouse images to obtain the body contour and key point information of each mouse; Step S3, obtaining the identity ID information of each mouse according to the mouse image information and body contour information based on a multi-target tracking algorithm; Step S4, calculating the body surface temperature of each mouse according to the mouse's ID information, body contour information and the temperature information; Step S5, calculating the individual behavior and social behavior of each mouse according to the mouse's identity ID information and key point information; Step S6, displaying the body temperature information and behavior of each mouse on a display screen in real time; In step S3, the Strongsort multi-target tracking algorithm is specifically used to initialize a track of the detected new target and assign an identity ID number. In the subsequent frames, the detected target and the existing track are matched according to the similarity of appearance features and position features, so as to distinguish the same mouse between the previous and next frames of the video; The similarity of the appearance features is represented by the minimum cosine distance, and the specific process is as follows: (1) Use the convolutional neural network OsNet to extract the appearance features of the detection target. The extraction range is the circumscribed rectangular box of the detection target, and a 512-dimensional feature vector is obtained; (2) Calculate the minimum cosine distance between the feature vector of each detected target and the feature vector of each existing trajectory to form a minimum cosine distance matrix. The smaller the minimum cosine distance value, the greater the similarity of the appearance features of the two, and the more likely they are the same target. The position feature similarity is represented by Mahalanobis distance, and the specific process is as follows: (1) Use the Kalman filter algorithm to predict the state information and corresponding covariance matrix of the next frame based on the state information and covariance matrix of the previous frame of the existing trajectory; (2) Based on the state information of the existing trajectory and the corresponding covariance matrix as well as the position information of the detected target, the Mahalanobis distance between the existing trajectory and the detected target is calculated to form a Mahalanobis distance matrix. The smaller the Mahalanobis distance, the smaller the position difference between the two, and the more likely they are the same target. The minimum cosine distance matrix and the Mahalanobis distance matrix are added with a weight of 0.95:0.05 to form a cost matrix; the smaller the value in the cost matrix, the lower the cost, which means that the corresponding existing track and the detection target are more likely to be matched successfully; the cost matrix is ​​input into the Hungarian algorithm to match the existing track and the detection target to achieve the recognition of the same mouse before and after the video; The state information of the existing trajectory is represented by an 8-dimensional vector Indicates that the location information of the detection target is ,in Indicates the coordinates of the center point of the mouse target box, is the aspect ratio of the frame, is the height of the frame, represent The velocity in the image coordinate system; The center point coordinates of the mouse target frame Set to the geometric centroid of the mouse body segmentation contour area; the aspect ratio of the frame Set to 1; the height of the box The body width is adapted to the body width of the mouse. The body width is obtained by fitting the body segmentation contour of the mouse to an ellipse using an ellipse fitting algorithm based on the least squares method. The length of the minor axis of the fitted ellipse is express, and The calculation relationship is .

2. A multi-target mouse behavior online recognition and body temperature monitoring method according to claim 1, characterized in that: In the step S2, the mouse image captured by the infrared thermal imaging camera is converted from a multi-channel color image to a single-channel grayscale image, and irrelevant background areas are cropped; the body contour of each mouse is segmented using the instance segmentation network YOLOv8-seg to obtain the body contour area information of each mouse; the body key points of each mouse are detected using the YOLOv8-pose key point detection network, and the key points include the tip of the nose, the left ear, the right ear, the neck, the center of the body and the base of the tail.

3. A multi-target mouse behavior online recognition and body temperature monitoring method according to claim 1, characterized in that: In step S4, the temperature information is a two-dimensional matrix of the same size as the image, which represents the temperature value of each pixel in the image; the average temperature value of the pixel points within the mouse body contour is calculated from the mouse body contour information as the mouse surface temperature; the surface temperature of each mouse can be obtained from the mouse identity ID information; according to the change in surface temperature, when the mouse body temperature is abnormal, an alarm message is issued.

4. A multi-target mouse behavior online recognition and body temperature monitoring method according to claim 1, characterized in that: In the step S5, the movement characteristics and body shape characteristics of the mouse and the distance between the nose tip of the mouse and the diet trough and the water port are calculated according to the key point information of the mouse; The motion feature is the displacement of the key points of the mouse body between the previous and next frames of the video; The body shape features are the distances between key points on the mouse body and the angles between the lines connecting the key points, specifically: Body length, from the tip of the nose to the base of the tail Distance from nose tip to left ear Distance from nose tip to right ear Distance from left ear to neck Distance from right ear to neck Distance from neck to body center Distance from the center of the body to the base of the tail Head rotation angle, that is, the angle between the line from the center of the body to the neck and the line from the neck to the tip of the nose Nose tip angle, that is, the angle between the line between the left ear and the nose tip and the line between the right ear and the nose tip Body rotation angle, that is, the angle between the line from the base of the tail to the center of the body and the line from the center of the body to the neck According to the above movement characteristics, body characteristics and the distance between the mouse nose tip and the diet trough and water drinking port, the support vector machine SVM was used to discriminate the individual behavior of the mice, and the kernel function of the support vector machine was selected as the Gaussian radial basis function; the identifiable individual behaviors included: eating, drinking, standing, sniffing, grooming, standing still and walking; the distances between the key points of the bodies of multiple mice and the head orientations were calculated to discriminate the social behaviors between the mice, and the social behaviors included: face sniffing, body sniffing, tail sniffing, chasing, fighting, and approaching; the number of occurrences and duration of the behavior of each mouse were recorded.

Images