Small animal detection device based on holder double-camera technology and detection method thereof

Abstract

The invention relates to a small animal detection device based on a holder double-camera technology and a detection method thereof. The device comprises a holder, a 3D camera, a zoom camera and an edge calculation controller, the holder at least has two rotational degrees of freedom; the 3D camera and the zoom camera are installed on the holder side by side. The edge calculation controller is respectively connected with the holder, the 3D camera and the zoom camera, so that the rotation angle of the holder can be controlled, and whether a small animal exists or not can be judged according to information acquired by the 3D camera and the zoom camera; according to the invention, rough scanning is carried out through the 3D camera, whether abnormity exists is rapidly judged according to the morphology difference value, and the abnormal position is amplified and captured through the zoom camera, so that the to-be-detected target has a large proportion in the picture, interference of light and shadow changes and the like on a common visible light camera is avoided, misjudgment is reduced, identification is facilitated, and the detection accuracy is improved. The system has the advantages of high efficiency, clear snapshot, high yield and the like, and meanwhile, the pan-tilt expands the scanning range and can cover a large range.

Description

Technical Field

[0001] The present invention relates to the technical field of small animal detection, and in particular to a small animal detection device based on a pan-tilt dual-camera technology and a detection method thereof. Background Art

[0002] For some important facilities, such as substations, machine rooms, laboratories, etc., small animals such as mice, snakes, birds, etc., once entering, may damage cables, chew the secondary wires of terminal boxes, or climb inside equipment casings and equipment. These behaviors are very likely to cause accidents such as short circuits and insulation damage, posing a serious threat to the safe operation of equipment. Therefore, by detecting small animals and giving timely alarms, measures can be taken quickly to prevent small animals from entering, thereby ensuring the normal operation of equipment.

[0003] Traditional methods of detecting small animals mainly rely on cameras for dynamic monitoring or capturing, and use the naked eye or algorithms to identify the movements of small animals. Since small animals are small in size, they occupy a small proportion of the monitoring image, are small targets within the range, and are not easy to be identified. In addition, light, shadow, debris, wind, etc. can easily cause interference. In addition, the movement speed of small animals is faster than the rotation speed of the camera, so it is difficult for the camera to keep tracking and capture clearly, and it is easy to capture residual images, light and shadow, and body parts, resulting in a low detection rate for small animals. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a small animal detection device and a detection method based on pan-tilt dual-camera technology. For indoor environments with long corridors such as substations, a device using a guide pan-tilt combined with a 3D camera and a zoom camera is designed for small animal detection.

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a small animal detection device based on a pan-tilt dual-camera technology, comprising a pan-tilt, a 3D camera, a zoom camera and an edge computing controller;

[0006] The pan / tilt head has at least two rotational degrees of freedom;

[0007] The 3D camera and the zoom camera are installed side by side on the pan-tilt platform; the 3D camera is used for rough scanning and judging whether there is any abnormality according to the difference in appearance; the zoom camera zooms in on the abnormal position and takes photos for identification;

[0008] The edge computing controller is connected to the pan-tilt head, 3D camera and zoom camera respectively, which can not only control the rotation angle of the pan-tilt head, but also determine whether there are small animals based on the information collected by the 3D camera and the zoom camera.

[0009] Preferably, it also includes a horizontally placed linear module; the pan head is arranged on the driving end of the linear module.

[0010] Preferably, it also includes three angle adjustment mechanisms; the three angle adjustment mechanisms are arranged on the back of the linear module and are placed in a triangle;

[0011] Each of the angle adjustment mechanisms includes a base plate, a movable plate placed parallel to the base plate, a screw rod vertically arranged on the base plate and passing through the movable plate, a spring sleeved on the screw rod and located between the base plate and the movable plate, a nut arranged on the outside of the movable plate and threadedly connected to the screw rod, and a guide rod vertically arranged on the base plate for guiding the movable plate; the screw rod and the guide rod are respectively located at both ends of the base plate; the movable plate is divided into an installation area and an adjustment area, wherein the installation area is connected to the linear module, and the screw rod passes through the adjustment area.

[0012] The present invention also discloses a small animal detection method based on a pan-tilt dual-camera technology, comprising the following steps:

[0013] S1: The 3D camera is driven by the PTZ to rotate along a preset trajectory to collect background shape data in the monitoring area;

[0014] S2: Entering the formal scanning and testing process, every time the gimbal rotates to a certain angle, the 3D camera collects the topography data, performs a difference operation with the background topography data at the corresponding position, and compares and determines whether there is a target interval with a significant difference value;

[0015] S3: If yes, adjust the magnification of the zoom camera according to the size of the corresponding target interval to be inspected;

[0016] S4: If the target interval to be inspected is not within the field of view of the current zoom camera at the magnification, the pan / tilt is adjusted to a corresponding position so that the target interval to be inspected falls within the field of view of the zoom camera;

[0017] S5: The edge computing controller then identifies the photo taken by the zoom camera and uses an algorithm to identify it. If it is determined that the target to be inspected is a small animal, an alarm is uploaded;

[0018] S6: The PTZ returns to the corresponding position and continues to rotate to the next position for scanning.

[0019] Preferably, in S1, when the 3D camera collects background morphology data in the monitoring area, the data unit is a three-dimensional point cloud data structure, represented by a three-dimensional coordinate array, which records the point cloud data of the entire background area; the background data is spatially divided, and the entire background area is divided into a three-dimensional grid; for each grid unit, the statistical information of the points falling into the unit, such as the average height and the number of points, is recorded.

[0020] Preferably, in S2, when an animal breaks into the background environment, the 3D camera is used again to collect the morphological data of the current scene to obtain new point cloud data; all three-dimensional grids are traversed, and for each grid cell, the corresponding cell statistical information is compared, and the height difference or point number difference is calculated. According to the set threshold, it is judged whether there is an animal in the grid cell, and the position of the animal is estimated by calculating the center position of the above cell.

[0021] Preferably, the center positions of all grid cells determined to contain animal parts are weighted averaged to obtain a more accurate position of the animal.

[0022] Preferably, the position of the animal in the 3D camera coordinate system is calculated and converted into the angular coordinates of the pan-tilt platform, thereby adjusting the corresponding position of the pan-tilt platform so that the target interval to be inspected falls within the field of view of the zoom camera.

[0023] Preferably, S5 also includes data collection and processing.

[0024] Carry out data collection in a variety of scenarios to build a high-quality small animal intrusion detection dataset; collect pictures containing small animal invasions as positive samples, and collect pictures without small animal invasions as negative samples. The number of positive samples of small animal targets should be no less than 2,000, and cover small animal targets of different species and behavioral characteristics; at the same time, in order to ensure the balance of data distribution, the number of positive and negative samples needs to be controlled so that the proportion of each type of sample is relatively balanced, thereby avoiding data bias affecting model performance;

[0025] After completing data collection, the data is divided into a training set and a test set in a ratio of 9:1. The training set is used for model training, and the test set is used for performance verification. The data division process should ensure the representativeness of different scenarios, sample types and features as much as possible.

[0026] Preferably, in S5, the algorithm recognition uses a deep learning network for image feature extraction and model learning, and uses a fully annotated training data set to perform multiple rounds of model training, and optimizes model parameters through repeated iterations;

[0027] The entire model training process includes the following steps:

[0028] S51: Feature extraction and network input,

[0029] Input the collected images into the deep learning network to extract multi-level feature information in the image, where the features include both the underlying detailed texture information and the high-level semantic features;

[0030] S52: Feature fusion and target prediction,

[0031] The extracted features are subjected to the interaction and fusion of upper and lower layer features through the FPN and PAN structures, and then the detection head uses the convolution operation to generate the target prediction box and the corresponding classification confidence score;

[0032] S53: Result optimization,

[0033] Non-maximum suppression is used to filter the detection results, remove candidate boxes with low confidence or large overlap, and finally output the location and category information of the small animal target.

[0034] Due to the application of the above technical solution, the present invention has the following advantages compared with the prior art:

[0035] The present invention first performs a rough scan with a 3D camera, quickly determines whether there is any abnormality based on the morphology difference, and then zooms in and captures the abnormal position with a zoom camera, so that the target to be inspected occupies a larger area in the picture. It not only avoids the interference of light and shadow changes on ordinary visible light cameras, reduces misjudgment, and facilitates identification. It has the advantages of high efficiency, clear capture, and high output rate. At the same time, the pan / tilt expands the scanning range and can cover a larger range. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The technical solution of the present invention is further described below in conjunction with the accompanying drawings:

[0037] Attached Figure 1 It is a structural schematic diagram of the small animal detection device based on the pan-tilt dual-camera technology of the present invention;

[0038] Attached Figure 2 This is a schematic diagram of the structure of the small animal detection device based on the pan-tilt dual-camera technology of the present invention after a linear module is added;

[0039] Attached Figure 3 It is a structural schematic diagram of the angle adjustment mechanism in the small animal detection device based on the pan-tilt dual-camera technology of the present invention;

[0040] Attached Figure 4 It is a flow chart of the small animal detection method based on the pan-tilt dual-camera technology of the present invention;

[0041] Attached Figure 5 A schematic diagram of a scanning channel of the small animal detection method based on the pan-tilt dual-camera technology of the present invention;

[0042] Attached Figure 6 A schematic diagram of the field of view of a 3D camera and a zoom camera in the small animal detection method based on the pan-tilt dual-camera technology of the present invention;

[0043] Attached Figure 7 Schematic diagram of the training process stages in the small animal detection method based on the pan-tilt dual camera technology described in the present invention.

[0044] Among them: 1. gimbal; 11. tangential axis; 12. normal axis; 2. 3D camera; 21. 3D camera field of view; 3. zoom camera; 31. zoom camera magnified field of view; 4. linear module; 5. angle adjustment mechanism; 51. bottom plate; 52. movable plate; 521. installation area; 522. adjustment area; 53. screw; 54. main spring; 55. nut; 56. guide rod; 57. auxiliary spring; 6. target to be inspected. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0046] Attached Figure 1-3 The small animal detection device based on the pan-tilt dual-camera technology of the present invention comprises a pan-tilt 1, a 3D camera 2, a zoom camera 3 and an edge computing controller;

[0047] The pan / tilt head 1 has two rotational degrees of freedom;

[0048] The 3D camera 2 and the zoom camera 3 are installed side by side on the pan / tilt 1; the 3D camera 2 is used for rough scanning and judging whether there is any abnormality according to the difference in appearance; the zoom camera 3 zooms in on the abnormal position and takes photos for identification;

[0049] The edge computing controller is connected to the pan-tilt head 1, the 3D camera 2 and the zoom camera 3 respectively, and can not only control the rotation angle of the pan-tilt head 1, but also determine whether there are small animals based on the information collected by the 3D camera 2 and the zoom camera 3.

[0050] In this embodiment, the two rotation axes of the pan / tilt head 1 are the tangential axis 11 and the normal axis 12. The rotation of the tangential axis 11 makes the field of view of the camera irregularly fan-shaped, and the entire corridor can be covered by adjusting the focal length; the rotation of the normal axis 12 expands the field of view of the single tangential axis 11 rotation, so that the field of view of the camera changes from a fan-shaped field of view to a full space.

[0051] Due to the parallel connection form and parallel distance of the 3D camera 2 and the zoom camera 3, and the physical limitations of their own volumes and the physical limitations of the shell processing technology, the images of the two cameras may not exactly contain the covered object. By linking the tangential axis 11 and the normal axis 12, the object can be photographed at a specific angle to achieve excellent image quality.

[0052] Furthermore, it also includes a horizontally placed linear module 4; the pan / tilt head 1 is arranged on the driving end of the linear module 4, so as to increase the search range of the pan / tilt head 1 and increase the shooting angle; if arranged at a corner, L-shaped or even U-shaped full-angle monitoring can be achieved, and from a physical level, the search length of the corridor is increased.

[0053] Furthermore, it also includes three angle adjustment mechanisms 5; the three angle adjustment mechanisms 5 are arranged on the back of the linear module 4, so that the detection device is suitable for various installation scenarios, and the search angle can be increased by adjusting the angle.

[0054] In this embodiment, the three angle adjustment mechanisms 5 are placed in a triangle, two of which are located at the upper ends of the linear module 4, and the remaining one is located in the lower middle part. The three points determine a plane to produce an angle adjustment effect; of course, if the linear module 4 is longer, the angle adjustment mechanism 5 is not limited to three.

[0055] Furthermore, each of the angle adjustment mechanisms 5 includes a base plate 51, a movable plate 52 placed parallel to the base plate 51, a screw 53 vertically arranged on the base plate 51 and passing through the movable plate 52, a main spring 54 sleeved on the screw 53 and located between the base plate 51 and the movable plate 52, a nut 55 arranged on the outer side of the movable plate 52 and threadedly connected to the screw 53, and a guide rod 56 vertically arranged on the base plate 51 for guiding the movable plate 52; the screw 53 and the guide rod 56 are respectively located at both ends of the base plate 51; the movable plate 52 is divided into an installation area 521 and an adjustment area 5, wherein the installation area 521 is connected to the linear module 4, and the screw 53 passes through the adjustment area 5 for easy adjustment; during installation: fix the base plate 51 on the wall, adjust the position of the movable plate 52 by rotating the nut 55, so that the movable plate 52 can achieve a slight movement under the bidirectional action of the main spring 54 and the nut 55, wherein the guide rod 56 guides the left and right, making the moving process smoother.

[0056] In this embodiment, in order to facilitate the installation of the linear module 4, the guide rod 56 is fixed on the movable plate 52, and can be moved through the bottom plate 51 and extended into the wall. At the same time, in order to ensure stable shooting, the spring can be selected with a large elastic coefficient, depending on the overall weight brought by the length of the guide rail.

[0057] Furthermore, an auxiliary spring 57 is sleeved on the guide rod 56 between the bottom plate 51 and the movable plate 52, which further plays a supporting role, so that both ends of the movable plate 52 are supported after installation, making it more stable. The elastic force of the auxiliary spring 57 is smaller than the elastic force of the main spring 54, thereby reducing the skewness of the movable plate 52.

[0058] Furthermore, the movable plate 52 is in a stepped shape, wherein the adjustment area 52 is higher than the installation area 521, which can play a positioning role during installation.

[0059] like Figure 4-6 As shown, the present invention also discloses a small animal detection method based on the pan-tilt dual-camera technology, comprising the following steps:

[0060] S1: Drive the 3D camera 2 to rotate once according to the preset trajectory through the pan / tilt 1 to collect background shape data in the monitoring area;

[0061] S2: Entering the formal scanning and testing process, every time the pan / tilt 1 rotates to a certain angle, the 3D camera 2 collects the topography data, performs a difference operation with the background topography data at the corresponding position, and compares and determines whether there is a target interval to be inspected with a significant difference value;

[0062] S3: If yes, adjust the magnification of the zoom camera 3 according to the size of the corresponding target interval to be inspected;

[0063] S4: If the target interval to be inspected is not within the current zoom camera magnification field of view 31 at the magnification ratio, the pan / tilt 1 is adjusted to a corresponding position so that the target interval to be inspected falls within the zoom camera magnification field of view 31;

[0064] S5: The edge computing controller then identifies the photo taken by the zoom camera 3 and performs identification through an algorithm. If it is determined that the target 6 to be inspected is a small animal, an alarm is uploaded;

[0065] S6: PTZ 1 returns to the corresponding position and continues to rotate to the next position for scanning.

[0066] Furthermore, in S1, we first obtain the data returned by the 3D camera. The data unit is a three-dimensional point cloud data structure. We use a three-dimensional coordinate array, such as [(x1, y1, z1), (x2, y2, z2), ...]), to record the point cloud data Background_PointCloud of the entire background area. We perform spatial division on the background data, and divide the entire background area into a three-dimensional grid (assuming that the grid unit size is grid_size). For each grid unit, we can record some statistical information of the points falling into the unit, such as the average height, number of points, etc. This step can be achieved by traversing the points in Background_PointCloud, calculating the grid units to which they belong, and updating the statistical information of the corresponding units.

[0067] Furthermore, in S2, when the animal breaks into the background environment, the 3D camera is used again to collect the morphological data of the current scene to obtain new point cloud data Current_PointCloud; all three-dimensional grid cells are traversed; for each grid cell, the cell statistics corresponding to Current_PointCloud and Background_PointCloud are compared; the height difference (if height is a key feature) or the point number difference can be calculated.

[0068] Assume that Current_Height[i] is the average height of the i-th grid cell in Current_PointCloud, and Background_Height[i] is the average height of the i-th grid cell in Background_PointCloud; calculate the height difference Height_Difference[i] = Current_Height[i] - Background_Height[i].

[0069] The set threshold is used to determine whether there is an animal in the grid cell; if |Height_Difference[i]|>threshold, it is considered that the cell may contain animal parts; the position of the animal can be estimated by calculating the center position of these cells. Assuming that the three-dimensional coordinate range of a grid cell is (x_min, y_min, z_min) to (x_max, y_max, z_max), the center position of the cell Center[i] = ((x_max+x_min) / 2, (y_max+y_min) / 2, (z_max+z_min) / 2).

[0070] The center positions of all grid cells that are judged to contain animal parts can be weighted averaged (the weight can be determined based on the size of the difference, etc.) to obtain a more accurate position estimate of the animal; if there are n grid cells that are judged to contain animal parts, their center positions are Center[1], Center[2], ..., Center[n], and their differences are Height_Difference[1], Height_Difference[2], ..., Height_Difference[n], then the animal position Animal_Position = (sum(Height_Difference[i]*Center[i]) / sum(Height_Difference[i])).

[0071] Furthermore, in S4, the position of the animal in the 3D camera coordinate system is calculated by an algorithm; we need to convert this position information into the angular coordinates of the gimbal 1; the mapping relationship from the 3D camera coordinate system to the gimbal 1 coordinate system;

[0072] The position and direction of the 3D camera and the gimbal 1 are fixed and can be calculated through geometric relationships. Assume that the 3D camera is located at the origin (0, 0, 0) and the horizontal angle range of the gimbal 1 is [α m i n , α max ], the vertical angle range is [β mi n ,β max ]; The corresponding angle can be calculated based on the relative relationship between the animal's position and the camera;

[0073] Assume the animal is at (x, y, z), the horizontal angle α = arctan(y / x), and the vertical angle Then, the angle of the pan / tilt 1 is controlled by PTZ and mapped to the target range for image capture and recognition.

[0074] Preferably, S5 also includes data collection and processing.

[0075] Data collection is carried out in a variety of scenarios to build a high-quality small animal intrusion detection dataset; specifically, pictures containing small animal invasions are collected as positive samples, and pictures without small animal invasions are collected as negative samples to ensure the diversity and authenticity of the data; to improve the generalization ability of the model, the number of positive samples of small animal targets should be no less than 2,000, and cover small animal targets of different types and behavioral characteristics; at the same time, to ensure the balance of data distribution, the number of positive and negative samples needs to be controlled so that the proportion of each type of sample is relatively balanced, thereby avoiding data bias from affecting model performance.

[0076] After completing data collection, the data is divided into a training set and a test set in a ratio of 9:1. The training set is used for model training and the test set is used for performance verification. During the data division process, the representativeness of different scenarios, sample types, and features should be ensured as much as possible to improve the applicability and robustness of the model in practical applications.

[0077] Furthermore, in S5, the algorithm recognition is based on the PyTorch open source framework, and a deep learning network dedicated to image feature extraction and model learning is built. The fully labeled training data set is used for multiple rounds of model training, and the model parameters are optimized through repeated iterations;

[0078] Before training starts, an adaptive anchor box generation mechanism is used to automatically adjust the anchor box size according to the distribution characteristics of the target sample, thereby improving the accuracy of target box prediction; in addition, a variety of data enhancement techniques are introduced, including but not limited to random cropping, brightness and color changes, image rotation, scaling, etc., to enhance the model's adaptability to diverse inputs and improve its generalization performance.

[0079] like Figure 7 As shown in the figure, the entire model training process includes the following steps:

[0080] S51: Feature extraction and network input

[0081] First, the collected images are input into the deep learning network. The backbone network is based on the CSPDarknet53 architecture and is responsible for extracting multi-level feature information in the image. These features contain both underlying detailed texture information and high-level semantic features, laying the foundation for target detection.

[0082] S52: Feature Fusion and Target Prediction

[0083] The extracted features are used to interact and fuse the upper and lower layer features through the FPN (Feature Pyramid Network) and PAN (Path Aggregation Network) structures. Subsequently, the detection head uses convolution operations to generate the target prediction box and the corresponding classification confidence score.

[0084] S53: Optimized results

[0085] Non-maximum suppression (NMS) is used to filter the detection results, remove candidate boxes with low confidence or large overlap, and finally output the location (such as bounding box coordinates) and category information (such as a certain animal type) of the small animal target.

[0086] During the training process, observe the downward trend of the loss value (loss) and the change curve of the average precision (mAP) in real time; terminate the training when the model training curve tends to be stable, and select the model weight file with the best performance based on the mAP value of the validation set.

[0087] Subsequently, the selected model is evaluated for inference in the test set, and performance indicators such as mAP, precision, recall, and F1 score are output. A confusion matrix is ​​drawn to analyze the distribution of false positives and false negatives of the model and to find potential problems.

[0088] Finally, the model is converted to adapt to actual deployment requirements. For example, the model format is converted into a hardware-adaptive file format to optimize its reasoning efficiency, and the model is deployed on the target hardware device to provide support for subsequent practical applications.

[0089] The complete design of this process ensures efficient training and performance optimization of the model, and provides a reliable technical solution for small animal target detection in actual scenarios.

[0090] The above are only specific application examples of the present invention and do not constitute any limitation on the protection scope of the present invention. Any technical solution formed by equivalent transformation or equivalent replacement shall fall within the protection scope of the present invention.

Claims

1. A small animal detection device based on pan-tilt dual-camera technology, characterized in that: Includes gimbal, 3D camera, zoom camera and edge computing controller; The pan / tilt head has at least two rotational degrees of freedom; The 3D camera and the zoom camera are installed side by side on the pan-tilt platform; the 3D camera is used for rough scanning and judging whether there is any abnormality according to the difference in appearance; the zoom camera zooms in on the abnormal position and takes photos for identification; The edge computing controller is connected to the pan-tilt head, 3D camera and zoom camera respectively, which can not only control the rotation angle of the pan-tilt head, but also determine whether there are small animals based on the information collected by the 3D camera and the zoom camera.

2. The small animal detection device based on the pan-tilt dual-camera technology according to claim 1 is characterized in that: It also includes a horizontally placed linear module; the pan head is arranged on the driving end of the linear module.

3. The small animal detection device based on the pan-tilt dual-camera technology according to claim 2 is characterized in that: It also includes three angle adjustment mechanisms; the three angle adjustment mechanisms are arranged on the back of the linear module and are placed in a triangle shape; Each of the angle adjustment mechanisms includes a base plate, a movable plate placed parallel to the base plate, a screw rod vertically arranged on the base plate and passing through the movable plate, a spring sleeved on the screw rod and located between the base plate and the movable plate, a nut arranged on the outside of the movable plate and threadedly connected to the screw rod, and a guide rod vertically arranged on the base plate for guiding the movable plate; the screw rod and the guide rod are respectively located at both ends of the base plate; the movable plate is divided into an installation area and an adjustment area, wherein the installation area is connected to the linear module, and the screw rod passes through the adjustment area.

4. The small animal detection method based on the pan-tilt dual camera technology according to claim 1 is characterized in that: The following steps are involved: S1: The 3D camera is driven by the PTZ to rotate along a preset trajectory to collect background shape data in the monitoring area; S2: Entering the formal scanning and testing process, every time the gimbal rotates to a certain angle, the 3D camera collects the topography data, performs a difference operation with the background topography data at the corresponding position, and compares and determines whether there is a target interval with a significant difference value; S3: If yes, adjust the magnification of the zoom camera according to the size of the corresponding target interval to be inspected; S4: If the target interval to be inspected is not within the field of view of the current zoom camera at the magnification, the pan / tilt is adjusted to a corresponding position so that the target interval to be inspected falls within the field of view of the zoom camera; S5: The edge computing controller then identifies the photo taken by the zoom camera and uses an algorithm to identify it. If it is determined that the target to be inspected is a small animal, an alarm is uploaded; S6: The PTZ returns to the corresponding position and continues to rotate to the next position for scanning.

5. The small animal detection method based on the PTZ dual camera technology according to claim 4 is characterized in that: In S1, when the 3D camera collects background shape data in the monitoring area, the data unit is a three-dimensional point cloud data structure represented by a three-dimensional coordinate array, which records the point cloud data of the entire background area; The background data is spatially divided, and the entire background area is divided into a three-dimensional grid; for each grid cell, the statistical information of the points falling into the cell is recorded, such as the average height and the number of points.

6. The small animal detection method based on the PTZ dual camera technology according to claim 5 is characterized in that: In S2, when an animal breaks into the background environment, the 3D camera is used again to collect the morphological data of the current scene to obtain new point cloud data; all three-dimensional grids are traversed, and for each grid cell, the corresponding cell statistical information is compared, and the height difference or point number difference is calculated. According to the set threshold, it is judged whether there is an animal in the grid cell, and the position of the animal is estimated by calculating the center position of the above cell.

7. The small animal detection method based on the PTZ dual camera technology according to claim 6 is characterized in that: The center positions of all grid cells judged to contain animal parts are weighted averaged to obtain a more accurate position of the animal.

8. The small animal detection method based on the PTZ dual camera technology according to claim 7 is characterized in that: In S4, the position of the animal in the 3D camera coordinate system is calculated and converted into the angular coordinates of the pan-tilt platform, thereby adjusting the corresponding position of the pan-tilt platform so that the target interval to be inspected falls within the field of view of the zoom camera.

9. The small animal detection method based on the PTZ dual camera technology according to claim 4 is characterized in that: S5 also includes data collection and processing. Carry out data collection in a variety of scenarios to build a high-quality small animal intrusion detection dataset; collect pictures containing small animal invasions as positive samples, and collect pictures without small animal invasions as negative samples. The number of positive samples of small animal targets should be no less than 2,000, and cover small animal targets of different species and behavioral characteristics; at the same time, in order to ensure the balance of data distribution, the number of positive and negative samples needs to be controlled so that the proportion of each type of sample is relatively balanced, thereby avoiding data bias affecting model performance; After completing data collection, the data is divided into a training set and a test set in a ratio of 9:

1. The training set is used for model training, and the test set is used for performance verification. The data division process should ensure the representativeness of different scenarios, sample types and features as much as possible.

10. The small animal detection method based on the PTZ dual camera technology according to claim 1, characterized in that: In S5, the algorithm recognition uses a deep learning network for image feature extraction and model learning, and uses a fully annotated training data set for multiple rounds of model training to optimize model parameters through repeated iterations; The entire model training process includes the following steps: S51: Feature extraction and network input, Input the collected images into the deep learning network to extract multi-level feature information in the image, where the features include both the underlying detailed texture information and the high-level semantic features; S52: Feature fusion and target prediction, The extracted features are subjected to the interaction and fusion of upper and lower layer features through the FPN and PAN structures, and then the detection head uses the convolution operation to generate the target prediction box and the corresponding classification confidence score; S53: Result optimization, Non-maximum suppression is used to filter the detection results, remove candidate boxes with low confidence or large overlap, and finally output the location and category information of the small animal target.

Images