Mouse detection method and system based on multi-frame judgment and multi-feature matching

Abstract

The present invention discloses a rat detection method and system based on multi-frame judgment and multi-feature matching. The method includes: obtaining monitoring images of a target area of ​​several frames for target detection. If a rat is detected in a certain frame, then obtaining M consecutive monitoring images after the frame; judging whether a rat is detected in each frame of the M monitoring images, if not, then determining that there is no rat in the current target area, if yes, then proceeding to the next step; judging whether the intersection-and-union ratios of the candidate detection frames of the rat target in all two adjacent frames are greater than 0.3, if not, then determining that there is no rat in the current target area, if yes, then proceeding to the next step; judging whether the intersection-and-union ratios of the candidate detection frames of the rat target in all two adjacent frames are less than 0.95, if yes, then determining that there is a rat in the current target area. The present invention combines the dynamic information of multi-frame images with the rat feature matching technology to effectively improve the accuracy of rat detection in complex scenes, especially in the application of grayscale images at night.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a mouse detection method and system based on multi-frame determination and multi-feature matching. Background Art

[0002] In important environments such as food processing, storage and kitchens, preventing rat intrusion is crucial to maintaining food safety, and timely and effective detection is essential. However, existing rat detection technologies face multiple challenges, especially in low-light and complex background environments.

[0003] At present, object detection technology based on computer vision and artificial intelligence has been applied in the field of rat detection. Usually, object detection models such as convolutional neural networks (CNN) are used to process surveillance images obtained by cameras to identify rats. However, these methods still have the following shortcomings:

[0004] 1. Grayscale image detection problem: At night, monitoring equipment usually uses infrared mode to obtain grayscale images. Due to the gray nature of mouse hair, the contrast in grayscale images is low, resulting in a high missed detection rate.

[0005] 2. False detection problem in complex environments: The size and color of mice in the environment are similar to some debris. The neural network model can easily mistake similar static objects for mice in complex scenes, especially in complex backgrounds and insufficient light, which can easily generate a large number of false alarms and reduce the reliability of the system.

[0006] 3. Insufficient detection accuracy of dynamic targets: The existing technology only relies on single-frame images for target detection, which makes it difficult to determine the movement status of mice, resulting in low accuracy in determining dynamic targets. Especially under various environmental interferences, information such as the movement trajectory and position changes of mice is not fully utilized, increasing the probability of missed detection and false detection. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the present invention aims to provide a rat detection method based on multi-frame judgment and multi-feature matching, which combines the dynamic information of multi-frame images with rat feature matching technology, effectively improving the accuracy of rat detection in complex scenes, especially in the application of night grayscale images.

[0008] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention includes:

[0009] The mouse detection method based on multi-frame determination and multi-feature matching includes the following steps:

[0010] S1. Obtain monitoring images of the target area in several frames, convert them into grayscale images, and use the trained mouse target monitoring model to perform target detection. If a mouse is detected in a certain frame, obtain M consecutive monitoring images after the frame.

[0011] S2, determining whether a mouse is detected in each of the M frames of monitoring images, if not, determining that there is no mouse in the current target area, if yes, proceeding to the next step;

[0012] S3, determine whether the intersection-over-union ratios of the candidate detection frames of the mouse target in all two adjacent frames are greater than 0.3. If not, it is determined that there is no mouse in the current target area. If so, proceed to the next step;

[0013] S4. Determine whether the intersection-over-union ratios of the candidate detection frames of the mouse target in all two adjacent frames are less than 0.95. If so, it is determined that there is a mouse in the current target area.

[0014] Preferably, step S4 further includes:

[0015] If not, extract the mouse features in each frame and calculate the average feature vector;

[0016] The similarity between the average feature vector and the reference feature vector is calculated. If the similarity is greater than 0.8, it is determined that there are mice in the current target area. If not, it is determined that there are no mice in the current target area. The reference feature vector includes several feature vectors extracted from the mouse image.

[0017] Preferably, in step S1, the confidence threshold for target detection using the trained mouse target monitoring model is 0.15-0.2.

[0018] Preferably, the reference feature vectors include: a group of mouse perspective feature vectors and a group of mouse part feature vectors; the mouse perspective feature vectors include feature vectors of the front, side and back of the mouse; the mouse part feature vectors include feature vectors of the mouse head, mouse body and complete mouse.

[0019] Preferably, the method for calculating the similarity between the average feature vector and the reference feature vector comprises:

[0020] The cosine similarity between the average feature vector and the mouse view feature vector and the mouse part feature vector is calculated respectively. If the cosine similarity between the average feature vector and at least one element feature vector in each group of reference feature vectors is greater than 0.8, it is determined that there is a mouse in the current target area.

[0021] The present invention also provides a mouse detection system based on multi-frame determination and multi-feature matching, comprising:

[0022] An image acquisition module is used to acquire monitoring images of a target area in several frames and convert them into grayscale images;

[0023] The target detection module has a built-in trained mouse target monitoring model, which is connected to the image acquisition module and is used to perform target detection on the grayscale image. If a mouse is detected in a certain frame, M consecutive monitoring images after the frame are acquired to determine whether a mouse is detected in each of the M monitoring images. If not, it is determined that there is no mouse in the current target area. If so, the M monitoring images are transmitted to the intersection-union ratio calculation module;

[0024] The intersection-and-union ratio calculation module is connected to the target detection module and is used to determine whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames in the M-frame monitoring image are greater than 0.3. If not, it is determined that there are no mice in the current target area; if so, it is determined whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames are less than 0.95. If so, it is determined that there are mice in the current target area.

[0025] Preferably, it also includes:

[0026] A reference feature vector module, which has several reference feature vectors extracted from mouse images;

[0027] A similarity calculation module, connected to the intersection-and-union ratio calculation module and the reference feature vector module respectively;

[0028] The intersection-and-union ratio calculation module determines whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames are less than 0.95, and if not, transmits the M frames of monitoring images to the similarity calculation module;

[0029] The similarity calculation module is used to extract the mouse features in each frame of the M-frame monitoring image, and calculate the average feature vector, and perform similarity calculation on the average feature vector and the reference feature vector. If the similarity is greater than 0.8, it is determined that there are mice in the current target area. If not, it is determined that there are no mice in the current target area.

[0030] Preferably, the confidence threshold of the mouse target monitoring model when performing target detection is 0.15-0.2.

[0031] Preferably, the reference feature vectors include: a group of mouse perspective feature vectors and a group of mouse part feature vectors; the mouse perspective feature vectors include feature vectors of the front, side and back of the mouse; the mouse part feature vectors include feature vectors of the mouse head, mouse body and complete mouse.

[0032] Preferably, the method in which the similarity calculation module calculates the similarity between the average feature vector and the reference feature vector comprises:

[0033] The cosine similarity between the average feature vector and the mouse view feature vector and the mouse part feature vector is calculated respectively. If the cosine similarity between the average feature vector and at least one element feature vector in each group of reference feature vectors is greater than 0.8, it is determined that there is a mouse in the current target area.

[0034] Beneficial Effects

[0035] 1. Reduce missed detection rate: This invention effectively reduces the possibility of missed detection of mice in grayscale images by lowering the confidence threshold (0.15-0.2) of the target detection model, especially solves the problem of difficulty in detecting mice due to low contrast in grayscale images, thereby significantly improving the sensitivity of detection. Then, a multi-frame judgment and feature matching mechanism is introduced, and the intersection over union (IoU) of the mouse candidate frames in multiple consecutive monitoring images is analyzed and feature matched, thus avoiding false detection caused by static interference in the environment.

[0036] 2. Multi-feature matching improves the recognition accuracy of static mice: For static targets with high IoU values ​​in multiple consecutive frames, the present invention extracts multiple perspectives and part features of the mouse, and uses cosine similarity calculation to match the mouse feature vector in the preset database to further confirm whether the static target is a mouse. This multi-feature matching technology effectively avoids the situation where static background objects are mistakenly identified as mice.

[0037] 3. Enhanced robustness in complex environments: The present invention can handle mouse detection problems in different scenes, lighting conditions, and background complexity. Whether the mouse is moving dynamically or staying statically, the method of the present invention can effectively deal with it, and through the combination of multi-frame judgment and feature matching, the robustness and adaptability of the system are greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic flow chart of a mouse detection method based on multi-frame determination and multi-feature matching provided in a preferred embodiment of the present invention;

[0039] Figure 2 A schematic diagram of the result of calculating the intersection-and-union ratio of candidate detection frames of a mouse target in two adjacent frames using a mouse detection method based on multi-frame determination and multi-feature matching provided in a preferred embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the structure of a mouse detection system based on multi-frame determination and multi-feature matching provided in another preferred embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described below in conjunction with the accompanying drawings. In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside", "outside" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention.

[0042] Embodiment 1

[0043] like Figure 1 As shown, this embodiment provides a mouse detection method based on multi-frame determination and multi-feature matching, comprising the steps of:

[0044] S1. Obtain monitoring images of the target area in several frames, convert them into grayscale images, and use the trained mouse target monitoring model to perform target detection. If a mouse is detected in a certain frame, obtain M consecutive monitoring images after the frame.

[0045] The target area refers to a specific area or place for rat monitoring, which is a specific part of the monitoring range covered by the camera. In the present invention, the target area can be an area in a key environment such as a food warehouse, kitchen, factory, etc., and the image of the area is monitored to determine whether there are rats.

[0046] Since mice have the habit of hiding during the day and coming out at night, night monitoring becomes particularly important. Night monitoring usually relies on infrared cameras, and the images generated by infrared cameras are grayscale images, lacking color information. Therefore, in order to ensure the detection consistency of the system in different environments during the day and night, the present invention takes into account the universal design, and directly converts the monitoring image into a grayscale image after acquiring it. This design not only avoids the difference between daytime color images and nighttime grayscale images, but also enables the system to maintain a consistent processing flow under any lighting conditions, reducing the fluctuation of detection accuracy caused by image mode switching.

[0047] Furthermore, color images are susceptible to interference from irrelevant factors such as background color and ambient light in certain complex environments. For example, a bright or colorful background may mislead the detection system and increase the possibility of false detection. The hair of mice usually presents gray or black features. Converting the image into a grayscale image helps to remove color interference and allows the detection system to focus more on core features such as the shape, outline, and brightness contrast of the target. This processing method not only improves the recognition accuracy of mouse targets, but also reduces the impact of irrelevant colors in the environment on detection, and enhances the robustness of the system in complex backgrounds. By uniformly processing images into grayscale images, the present invention achieves all-weather detection consistency, while improving detection accuracy and adaptability in a variety of scenarios.

[0048] Since the mouse's hair is gray, the contrast with the background in the grayscale image is low, which leads to certain limitations in the detection of mice by existing target detection algorithms (such as the YOLOv9 model based on convolutional neural networks), especially in the grayscale image environment, the missed detection rate of mice is high. Taking the YOLOv9 target detection model as an example, its default confidence threshold (thresh) is 0.25. Under this condition, although the detection accuracy of the model is high, the missed detection phenomenon is more serious due to the insufficient contrast between the mouse and the background in the grayscale image.

[0049] In order to improve the sensitivity of detection and reduce missed detections, the present invention, in some preferred embodiments, appropriately lowers the confidence threshold of the detection model and sets it to between 0.15 and 0.2. By lowering the confidence threshold, the system can capture more mouse target candidate frames and expand the detection range, which greatly reduces the missed detection rate of mice. However, this processing method also introduces a new problem, that is, while obtaining more candidate detection frames, the detection system may also generate a large number of non-mouse candidate frames, resulting in an increase in the false detection rate. In order to address this problem, the present invention designs a subsequent multi-step detection method.

[0050] It should be noted that if a mouse does exist in the monitoring area, it will be detected in multiple consecutive frames of images regardless of whether it is in a moving state or a stationary state. Therefore, the present invention introduces a multi-frame judgment mechanism, which is used to further verify the detection results by setting a threshold M for multi-frame judgment. Specifically, when a mouse target is detected in a certain frame, the subsequent M frames of continuous images will be subjected to target detection and analysis to ensure that the target is an actual mouse, rather than an accidentally appearing interference or a misdetected object. Through the multi-frame judgment mechanism, the present invention can effectively distinguish between a real mouse target and a non-mouse target that occasionally appears in a single frame image. In the continuous M frames of images, if a mouse target can be detected in each frame, and the intersection over union (IoU) between the detection frames meets certain movement characteristics, it can be confirmed that the target is a real mouse, especially during its movement. If a mouse is detected only in a certain frame, and continuous detection is not maintained in subsequent frames, or the intersection over union of the detection results shows that the target does not meet the movement characteristics of a mouse, the possibility that the target is a mouse can be ruled out. By setting the multi-frame determination threshold M, the present invention effectively reduces the possibility of occasional false detections, while ensuring that the actual mouse target can be robustly detected, regardless of whether it is in a moving or stationary state. This technical means reduces the false detection rate while also ensuring continuous tracking and high-precision detection of the target.

[0051] S2, determine whether mice are detected in each of the M frames of monitoring images. If not, it is determined that there are no mice in the current target area. If yes, proceed to the next step. Only when mice are detected in all the consecutive M frames of monitoring images, it is considered that mice may exist. Otherwise, it is considered that there are interference objects other than mice.

[0052] S3. Determine whether the intersection-over-union ratios of the candidate detection frames of the mouse target in all two adjacent frames are greater than 0.3. If not, it is determined that there is no mouse in the current target area. If so, proceed to the next step.

[0053] S4. Determine whether the intersection-over-union ratios of the candidate detection frames of the mouse target in all two adjacent frames are less than 0.95. If so, it is determined that there is a mouse in the current target area.

[0054] The technical logic of the above steps is as follows: Intersection over Union (IoU) is used to quantify the overlap between two detection frames and is an important indicator for measuring the position change of the target in consecutive frames. In the present invention, the IoU value is used to distinguish mice from other possible interferences.

[0055] When the IoU value is less than 0.3, it means that the overlapping area of ​​the two detection frames is small and the target has moved significantly between consecutive frames. For mice, this rapid and large movement does not conform to their usual movement characteristics, especially in confined spaces. It can be inferred that the target in this situation is more likely to be other high-speed interference objects, such as flying insects or fast-moving debris, so the possibility that the target is a mouse can be directly ruled out.

[0056] When the IoU value is between 0.3 and 0.95, such as Figure 2 As shown in the figure, it means that the target moves between consecutive frames, and the movement amplitude is moderate. This movement feature is more consistent with the natural activity mode of mice in the monitoring area, such as crawling or slowly crossing. Therefore, the target is more likely to be a moving mouse at this time, and should be further retained to continue the subsequent detection steps.

[0057] When the IoU value is greater than 0.95, it means that the detection box has almost no change between consecutive frames, and the target appears to be stationary. However, a stationary target may be a stationary mouse or other static interference (such as debris in the background). Therefore, for these targets with high IoU values, movement information alone is not enough to make an accurate judgment, and further feature matching steps are still required, such as comparison of multi-view features or part features of the mouse, to distinguish its true identity.

[0058] It should be noted that the setting of the intersection over union (IoU) thresholds, i.e., 0.3 and 0.95, is based on the historical analysis of the movement speed and state of mice in different possible scenarios. These thresholds reflect the typical behavior patterns and movement characteristics of mice in the monitoring environment. By analyzing the historical data of mouse behavior, the present invention effectively sets these IoU thresholds, so that the detection system can more accurately judge the movement state of the target, thereby improving the accuracy and reliability of mouse detection.

[0059] In some preferred embodiments, a method for determining when the IoU is greater than 0.95 is provided, and step S4 further includes:

[0060] If the intersection-and-union ratio of the mouse target candidate detection frames in all two adjacent frames is greater than 0.95, the mouse features in each frame are extracted and the average feature vector is calculated. Specifically, for M consecutive monitoring images containing at least M mouse target candidate detection frames, feature extraction is performed on each mouse target candidate detection frame. At this time, each feature is a one-dimensional vector with 1024 dimensions. The average feature vector is calculated for these M mouse features. The calculation formula is as follows:

[0061] ;

[0062] Where i is the eigenvector number, j is the element number in the eigenvector, represents the jth element of the i-th eigenvector, is the sum of the jth elements of the M eigenvectors, is the average eigenvector.

[0063] The similarity between the average feature vector and the reference feature vector is calculated. If the similarity is greater than 0.8, it is determined that there are mice in the current target area. If not, it is determined that there are no mice in the current target area. The reference feature vector includes several feature vectors extracted from the mouse image.

[0064] It should be noted that the reference feature vector refers to preset feature data used to compare the detection object (i.e. the detected mouse target) for similarity judgment. These feature vectors are mouse feature data extracted from multiple different angles and parts based on known mouse images, and are stored in the system database as an important basis for judging whether the target in the current monitoring image is a mouse. This is because the mouse target candidate detection frame may be different perspectives and different parts of the mouse. Matching these features can improve the accuracy. In some preferred embodiments, the specific composition of the reference feature vector is given, including a group of mouse perspective feature vectors and a group of mouse part feature vectors; the mouse perspective feature vector includes feature vectors of the front, side and back of the mouse; the mouse part feature vector includes feature vectors of the mouse head, mouse body and complete mouse. Further, the method for calculating the similarity between the average feature vector and the reference feature vector includes: respectively calculating the cosine similarity of the average feature vector with the mouse perspective feature vector and the mouse part feature vector, and if the cosine similarity between the average feature vector and at least one element feature vector in each group of reference feature vectors is greater than 0.8, it is determined that there is a mouse in the current target area. That is, each set of reference feature vectors must have at least one feature with a similarity greater than 0.8 with the average feature vector, and only then will it be judged that there is a mouse, otherwise it will be judged as other interferences.

[0065] Embodiment 2

[0066] like Figure 3 As shown, this embodiment provides a mouse detection system based on multi-frame determination and multi-feature matching, including:

[0067] An image acquisition module is used to acquire monitoring images of a target area in several frames and convert them into grayscale images;

[0068] The target detection module has a built-in trained mouse target monitoring model, which is connected to the image acquisition module and is used to perform target detection on the grayscale image. If a mouse is detected in a certain frame, M consecutive monitoring images after the frame are acquired to determine whether a mouse is detected in each of the M monitoring images. If not, it is determined that there is no mouse in the current target area. If so, the M monitoring images are transmitted to the intersection-union ratio calculation module;

[0069] The intersection-and-union ratio calculation module is connected to the target detection module and is used to determine whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames in the M-frame monitoring image are greater than 0.3. If not, it is determined that there are no mice in the current target area; if so, it is determined whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames are less than 0.95. If so, it is determined that there are mice in the current target area.

[0070] In some preferred embodiments, it also includes:

[0071] A reference feature vector module, which has several reference feature vectors extracted from mouse images;

[0072] A similarity calculation module, connected to the intersection-and-union ratio calculation module and the reference feature vector module respectively;

[0073] The intersection-and-union ratio calculation module determines whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames are less than 0.95, and if not, transmits the M frames of monitoring images to the similarity calculation module;

[0074] The similarity calculation module is used to extract the mouse features in each frame of the M-frame monitoring image, and calculate the average feature vector, and perform similarity calculation on the average feature vector and the reference feature vector. If the similarity is greater than 0.8, it is determined that there are mice in the current target area. If not, it is determined that there are no mice in the current target area.

[0075] In some preferred embodiments, the confidence threshold of the mouse target monitoring model when performing target detection is 0.15-0.2.

[0076] In some preferred embodiments, the reference feature vector includes: a group of mouse perspective feature vectors and a group of mouse part feature vectors; the mouse perspective feature vector includes feature vectors of the front, side and back of the mouse; the mouse part feature vector includes feature vectors of the mouse head, mouse body and complete mouse.

[0077] In some preferred embodiments, the method in which the similarity calculation module calculates the similarity between the average feature vector and the reference feature vector includes:

[0078] The cosine similarity between the average feature vector and the mouse view feature vector and the mouse part feature vector is calculated respectively. If the cosine similarity between the average feature vector and at least one element feature vector in each group of reference feature vectors is greater than 0.8, it is determined that there is a mouse in the current target area.

[0079] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. Mouse detection method based on multi-frame judgment and multi-feature matching, It is characterized in that Includes steps: S1. Obtain monitoring images of the target area in several frames, convert them into grayscale images, and use the trained mouse target monitoring model to perform target detection. If a mouse is detected in a certain frame, obtain M consecutive monitoring images after the frame. S2, determining whether a mouse is detected in each of the M frames of monitoring images, if not, determining that there is no mouse in the current target area, if yes, proceeding to the next step; S3, determine whether the intersection-over-union ratios of the candidate detection frames of the mouse target in all two adjacent frames are greater than 0.

3. If not, it is determined that there is no mouse in the current target area. If so, proceed to the next step; S4, determining whether the intersection-over-union ratios of the candidate detection frames of the mouse target in all two adjacent frames are less than 0.95, if so, it is determined that there is a mouse in the current target area; Step S4 also includes: If not, extract the mouse features in each frame and calculate the average feature vector; Calculate the similarity between the average feature vector and the reference feature vector. If the similarity is greater than 0.8, it is determined that there is a mouse in the current target area. If not, it is determined that there is no mouse in the current target area. The reference feature vector includes several feature vectors extracted from the mouse image. The reference feature vectors include: a group of mouse view feature vectors and a group of mouse part feature vectors; the mouse view feature vectors include feature vectors of the front, side and back of the mouse; the mouse part feature vectors include feature vectors of the mouse head, mouse body and complete mouse; The method for calculating the similarity between the average feature vector and the reference feature vector includes: The cosine similarity between the average feature vector and the mouse view feature vector and the mouse part feature vector is calculated respectively. If the cosine similarity between the average feature vector and at least one element feature vector in each group of reference feature vectors is greater than 0.8, it is determined that there is a mouse in the current target area.

2. The rat detection method based on multi-frame determination and multi-feature matching as claimed in claim 1, It is characterized in that In the step S1, the confidence threshold for target detection using the trained mouse target monitoring model is 0.15-0.

2.

3. Mouse detection system based on multi-frame judgment and multi-feature matching, It is characterized in that include: An image acquisition module is used to acquire monitoring images of a target area in several frames and convert them into grayscale images; The target detection module has a built-in trained mouse target monitoring model, which is connected to the image acquisition module and is used to perform target detection on the grayscale image. If a mouse is detected in a certain frame, M consecutive monitoring images after the frame are acquired to determine whether a mouse is detected in each of the M monitoring images. If not, it is determined that there is no mouse in the current target area. If so, the M monitoring images are transmitted to the intersection-union ratio calculation module; The intersection-and-union ratio calculation module is connected to the target detection module and is used to determine whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames in the M-frame monitoring image are greater than 0.

3. If not, it is determined that there are no mice in the current target area; if so, it is determined whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames are less than 0.

95. If so, it is determined that there are mice in the current target area; Also includes: A reference feature vector module, which has several reference feature vectors extracted from mouse images; A similarity calculation module, connected to the intersection-and-union ratio calculation module and the reference feature vector module respectively; The intersection-and-union ratio calculation module determines whether the intersection-and-union ratios of the mouse target detection frames in all two adjacent frames are less than 0.95, and if not, transmits the M frames of monitoring images to the similarity calculation module; The similarity calculation module is used to extract the mouse features in each frame of the M frames of monitoring images, and calculate the average feature vector, and perform similarity calculation on the average feature vector and the reference feature vector. If the similarity is greater than 0.8, it is determined that there are mice in the current target area, and if not, it is determined that there are no mice in the current target area; The reference feature vectors include: a group of mouse view feature vectors and a group of mouse part feature vectors; the mouse view feature vectors include feature vectors of the front, side and back of the mouse; the mouse part feature vectors include feature vectors of the mouse head, mouse body and complete mouse; The method in which the similarity calculation module calculates the similarity between the average feature vector and the reference feature vector includes: The cosine similarity between the average feature vector and the mouse view feature vector and the mouse part feature vector is calculated respectively. If the cosine similarity between the average feature vector and at least one element feature vector in each group of reference feature vectors is greater than 0.8, it is determined that there is a mouse in the current target area.

4. The rat detection system based on multi-frame determination and multi-feature matching as claimed in claim 3, Features: The confidence threshold of the mouse target monitoring model during target detection is 0.15-0.2.

Images