Method and system for detecting feeding of stray animals based on inspection robot

Abstract

The present application relates to the technical field of visual detection, and more particularly to a feeding stray animal behavior detection method and system based on a patrol robot, which first continuously collects initial images on a predetermined patrol path, then analyzes the initial images based on a preset real-time target detection model to obtain stray animal gathering features and human behavior features, then inputs the stray animal gathering features and human behavior features into a preset artificial intelligence analysis model to obtain a plurality of comprehensive judgment results, and finally drives the patrol robot to perform corresponding on-duty tasks according to the plurality of comprehensive judgment results. Unlike the prior art, the method effectively solves the problem of affecting the experience of passers-by caused by the low recognition accuracy of the existing public scene feeding detection technology.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology. More specifically, this invention relates to a method and system for detecting the feeding behavior of stray animals based on an inspection robot. Background Technology

[0002] The large gatherings of pigeons or stray cats and dogs in scenic spots, parks, or public squares have long been a significant factor affecting urban public health. The main reason for these animal gatherings is human feeding. This feeding behavior leads to the accumulation of animal feces, increasing the workload of sanitation workers; furthermore, it promotes the spread of viruses, bacteria, cryptococci, or toxoplasmosis.

[0003] To address the aforementioned issues and promote the development of smart scenic areas, existing technologies have already piloted patrol robots in some scenic areas to discourage tourists from feeding stray animals. For example, Chinese patent application CN112487926A analyzes human joint movements to identify whether tourists are feeding animals and issues an alert accordingly. However, this technology has the following drawback: Specifically, actions similar to feeding can also be considered "feeding stray animals." For instance, bending down to pick up a dropped object, tying shoelaces, falling forward, feeding one's own pet, or playful behavior between couples can all be interpreted as feeding stray animals, which negatively impacts the visitor experience.

[0004] Therefore, existing public feeding detection technologies suffer from low accuracy, which negatively impacts the experience of passersby. Summary of the Invention

[0005] To address the technical problem of low accuracy in existing public feeding detection technologies, which negatively impacts the visitor experience, this invention discloses a method and system for detecting the feeding behavior of stray animals based on an inspection robot.

[0006] In a first aspect, the present invention discloses a method for detecting the feeding behavior of stray animals based on an inspection robot, comprising:

[0007] Initial images are continuously collected along the predetermined inspection path; Based on a pre-defined real-time target detection model, the initial image is analyzed to obtain features of ownerless animal aggregation and human behavior. By inputting the characteristics of stray animal aggregation and human behavior into a pre-set artificial intelligence analysis model, a multi-frame comprehensive judgment result is obtained; Based on the comprehensive judgment results of multiple frames, the inspection robot is driven to perform the corresponding duty tasks.

[0008] Beneficial Effects: The method of this invention is suitable for deployment in inspection robots to perform tasks in residential areas, scenic spots, or other work areas. Regarding the feeding behavior of stray animals, this invention utilizes the collected images of public places such as residential areas and squares to perform dual analysis of stray animal aggregation characteristics and human behavioral characteristics on the inspection robot itself. Then, based on these characteristics, a refined artificial intelligence analysis task is performed, thereby filtering out false positives and effectively improving recognition accuracy and visitor experience.

[0009] Preferably, based on a preset real-time target detection model, the initial image is analyzed, including: Pick several ownerless animal frames from the initial image; The cluster centroid is calculated based on the center of several bounding boxes without an active action. Calculate the average Euclidean distance from the center of all ownerless animal frames to the centroid of aggregation; Based on the average Euclidean distance, the aggregation characteristics of ownerless animals are marked as either aggregated or non-aggregated.

[0010] Preferably, based on a preset real-time target detection model, the initial image is analyzed. The method of the present invention further includes: From the initial image, pick the human detection box that is closest to the ownerless animal box; Calibrate the key nodes and spatial coordinates of the human body detection frame; Key nodes are filtered based on the visual features or spatial coordinates of the initial image to obtain multiple credible key nodes; The spatial relationship between multiple trusted key nodes is calculated to obtain human behavioral characteristics that characterize the node spacing distance and trunk flexion and extension angle.

[0011] Preferably, human behavioral characteristics include the wrist-hip distance; by inputting the aggregation characteristics of stray animals and human behavioral characteristics into a preset artificial intelligence analysis model, multi-frame comprehensive judgment results are obtained, including: If the animal's aggregation characteristic is aggregation, and the trunk flexion-extension angle is greater than a preset first angle threshold, perform the following steps: Determine whether the wrist-hip interval is greater than the preset second interval distance. If so, obtain the current frame judgment result used to characterize the feeding state.

[0012] Preferably, after obtaining the judgment result of the current frame, the method of the present invention further includes: Convert multiple temporally consecutive initial images into RGBA four-channel images; The RGBA four-channel image and the judgment result of the current frame are input into the artificial intelligence analysis model to obtain a comprehensive judgment result of multiple frames.

[0013] Preferably, the artificial intelligence analysis model is configured as follows: Gaussian distribution sampling and downsampling are performed on the RGBA four-channel image to obtain the sampled image; Convolve the sampled image and the judgment result of the current frame to obtain temporal image features; Based on temporal image features and a pre-defined binary cross-entropy loss function, a multi-frame comprehensive judgment result is predicted.

[0014] Preferably, the real-time object detection model adopts a detection model based on YOLOV8; the real-time object detection model adopts the ReLU activation function; the real-time object detection model also replaces the upsample operator in the original YOLOV8 with a combination of operators including deconvolution operator, Batch Normalization operator, and ReLU activation operator.

[0015] Preferably, if the animal aggregation feature is aggregation, the human behavior feature detection step is performed; otherwise, the human behavior feature detection step for the current frame is terminated.

[0016] Preferably, the method of the present invention further includes: Pre-train the artificial intelligence analysis model.

[0017] Secondly, the present invention also discloses a feeding behavior detection system for stray animals based on an inspection robot, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the feeding behavior detection method for stray animals based on an inspection robot described in the first aspect is implemented.

[0018] The advantages of this invention, which differ from existing technologies, are as follows: Firstly, it effectively solves the problem that the low accuracy of existing public feeding detection technologies affects the experience of passersby.

[0019] Secondly, the target detection model has been improved, enhancing the accuracy and efficiency of feeding detection.

[0020] Third, configurational improvements were made to the artificial intelligence analysis model, which effectively enhanced the model's derivation capabilities and further improved the accuracy of feeding detection. Attached Figure Description

[0021] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart of the feeding behavior detection method for stray animals based on an inspection robot in Embodiment 1 of the present invention; Figure 2 This is a calibration diagram of key nodes within the human body detection frame in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the feeding behavior detection system for stray animals based on an inspection robot in Embodiment 2 of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] Example 1 like Figure 1 As shown, this embodiment discloses a method for detecting the feeding behavior of stray animals based on an inspection robot, including: S10: Continuously acquire initial images along the predetermined inspection path.

[0025] It should be explained that the method in this embodiment is used for an inspection robot, and the camera for acquiring the initial image is placed on the top of the inspection robot's head. This method continuously drives the inspection robot to patrol back and forth along the inspection path of public places such as residential areas and outdoor squares, acquiring initial images of these places in real time.

[0026] Furthermore, the initial image size captured by the inspection robot's camera is 1920. 1080, the initial image preprocessing method is to convert 1920... 1080-sized image downsampled to 768 The size is 448, and each pixel value is divided by 255.

[0027] S20: Based on the preset real-time target detection model, the initial image is analyzed to obtain the features of ownerless animal aggregation and human behavior.

[0028] In this embodiment, the real-time object detection model adopts a YOLOv8-based detection model. To accelerate the inference speed and accuracy of the object detection model, the method in this embodiment makes the following modifications to YOLOv8, specifically: Firstly, at the activation function level, the ReLU activation function was used instead of the original SiLU activation function in YOLOv8.

[0029] Secondly, the Upsample operator in YOLOV8 is replaced by a combination of operators including deconvolution, Batch Normalization, and ReLU activation.

[0030] Furthermore, for the analysis of ownerless animal aggregations, step S20 above includes: S211: Pick several ownerless animal frames from the initial image.

[0031] Specifically, a real-time object detection model is invoked to identify one or more ownerless bounding boxes, and then the center coordinates of each ownerless bounding box are calculated.

[0032] S212: The cluster centroid is calculated based on the center of several active action boxes.

[0033] Specifically, the formula for calculating the centroid of aggregation in the x-direction is as follows:

[0034] In the formula, Represents the x-coordinate of the centroid of aggregation; This represents the center coordinate of the i-th unclaimed animal frame in the x-direction. This indicates the number of unclaimed animal frames.

[0035] The specific formula for calculating the centroid of aggregation in the y-direction is as follows:

[0036] In the formula, Represents the y-coordinate of the centroid; This represents the center coordinate of the i-th unclaimed animal frame in the y-direction.

[0037] S213: Calculate the average Euclidean distance from the center of all ownerless animal frames to the centroid of the cluster.

[0038] The formula for calculating the average Euclidean distance is as follows:

[0039] In the formula, Let represent the average Euclidean distance of the i-th unclaimed animal frame.

[0040] S214: Based on the average Euclidean distance, mark the aggregation characteristics of ownerless animals as aggregated or non-aggregated.

[0041] Then, based on the above average Euclidean distance, we need to calculate the mean and variance of the Euclidean distances for all non-active objects. The specific calculation process for the mean is as follows:

[0042] The specific calculation process for variance is as follows:

[0043] Furthermore, to improve recognition accuracy, a reasonable threshold needs to be set. In this embodiment, when the mean is greater than 50 and the variance is less than 30, it can be considered that a gathering of ownerless animals caused by feeding has occurred. Otherwise, it is considered that no gathering has occurred.

[0044] Preferably, in one embodiment, to save computing resources and reduce computational load, if the animal aggregation feature is aggregated, the human behavior feature detection step is performed. Otherwise, the human behavior feature detection step for the current frame is terminated.

[0045] Furthermore, for the analysis of ownerless animal aggregations, step S20 above includes: S221: From the initial image, pick the human detection box that is closest to the ownerless animal box.

[0046] Select one or more human detection frames that are closest to the unclaimed animal frame.

[0047] Conversely, if only a group of stray animals is detected without a human detection frame, it is determined that there is no suspicious feeding of small animals, and the subsequent steps end directly.

[0048] S222: Calibrate the key nodes and spatial coordinates of the human body detection frame.

[0049] Combination Figure 2 As shown, in this embodiment, the Alphapose framework can be used to detect human key points. To increase the speed and accuracy of human key point detection and to facilitate model deployment and portability, the Alphapose framework is adapted as follows: First, the backbone network uses the Resnet18 network.

[0050] Secondly, the PixelShuffle operator in Alphapose remains unchanged during training, but is replaced by the equivalent Deconvolution operator during inference.

[0051] Regarding the number of keypoints detected by the Alphapose framework and the training dataset, the method in this embodiment makes the following adaptations: First, 15 key points are detected for each human body. Points numbered 1-13 are derived from annotations, while key points numbered 14 and 15 are not. Point number 14 is calculated by taking the midpoint between points numbered 3 and 2, and point number 15 is calculated by taking the midpoint between points numbered 9 and 8. This method ensures the strong robustness of key points 14 and 15 during the inference phase.

[0052] It should be added that the training dataset for the Alphapose framework mentioned above is a combination of four open-source datasets: COCO, Haple, Crowd, and MPII.

[0053] Furthermore, after calibrating the key human points in each human detection frame, two-dimensional spatial coordinates were assigned to them.

[0054] Finally, the data format for outputting 15 key points of the human body in the fall state is [15,x,y,conf], where 15 represents the number of key points, x and y represent the coordinate positions of the key points on the image, and conf represents the confidence level of the key points.

[0055] S223: Filter key nodes based on the visual features or spatial coordinates of the initial image to obtain multiple credible key nodes.

[0056] For example, Figure 2 The key nodes 14, 15, 3, 7, 2, and 6 in the image are assessed for confidence. Only key nodes with a confidence score greater than 0.5 are allowed subsequent logical checks; otherwise, they are considered abnormal or occluded and not considered further. Finally, for ease of compilation, these key nodes are named `neck`, `hip`, `left_shoulder`, `left_wrist`, `right_shoulder`, and `right_wrist`, respectively. The confidence score configuration is based on the visual features or spatial coordinates of the initial image to assign a score to each key node. For example, using key node 1 (the head node) as a baseline, and key node 3 as the left shoulder, the spatial distance between the two key nodes exceeds 35 cm, significantly exceeding the normal human morphological distribution. Therefore, key node 3 can be assigned a confidence score of 0.2. Correspondingly, other confidence score assignment logic also needs to be based on the normal human morphological distribution.

[0057] S224: Perform node spatial relationship calculations on multiple trusted key nodes to obtain human behavioral characteristics that characterize node spacing distances and trunk flexion-extension angles.

[0058] It should be noted that the above-mentioned human behavioral characteristics include the wrist-hip distance.

[0059] More specifically, during the execution of script programming, the programming configuration can be: Calculate the angle by the cosine of the line connecting nodes 14 (neck) and 15 (hip) and the perpendicular line, denoted as body_a. Calculate the angle by the cosine of the line connecting nodes 3 (left_shouder) and 7 (left_wrist) and the line connecting nodes 9 (neck) and 10 (hip), denoted as left_a. Similarly, calculate the angle by the cosine of the line connecting nodes 2 (right_shouder) and 6 (right_wrist) and the line connecting nodes 9 (neck) and 10 (hip), denoted as right_a. Calculate the distance from left_wrist to hip, denoted as left_d. Calculate the distance from right_wrist to hip, denoted as right_d.

[0060] The left_a (left bend angle), right_a (right bend angle), left_d, and right_d mentioned above are all human behavioral feature data.

[0061] S30: Input the characteristics of unclaimed animal aggregation and human behavior into a preset artificial intelligence analysis model to obtain a multi-frame comprehensive judgment result.

[0062] In a simplified embodiment, step S30 above includes: If the animal aggregation characteristic is aggregation and the trunk flexion-extension angle is greater than the preset first angle threshold, perform the following steps: determine whether the wrist-hip interval distance is greater than the preset second interval distance; if so, obtain the current frame judgment result used to characterize the feeding state.

[0063] Specifically, check if body_a is greater than the first angle threshold. If it is, proceed with the subsequent judgment; otherwise, the "suspicious feeding of stray animals" situation has not occurred. Check if at least one of right_d and left_d is greater than the second interval distance. If it is, the person is considered to have "suspiciously fed stray animals"; otherwise, the "suspicious feeding of stray animals" situation has not occurred.

[0064] In another embodiment with higher precision, step S30 above can also be: Check if body_a is greater than the first angle threshold. If so, proceed with further judgment; otherwise, the "suspicious feeding of stray animals" situation has not occurred. Check if at least one of left_a and right_a is greater than the second angle threshold. If so, proceed with further judgment; otherwise, the "suspicious feeding of stray animals" situation has not occurred. Finally, check if at least one of right_d and left_d is greater than the second interval distance. If so, the person is considered to have "suspiciously fed stray animals"; otherwise, the "suspicious feeding of stray animals" situation has not occurred.

[0065] It should be added that if at least one of the three people closest to the camera exhibits the "suspicious feeding of stray animals" state, then based on that frame, all video frames within 3 seconds before and after that frame are collected. Furthermore, a Region of Interest (ROI) is formed by expanding the center of the human detection box exhibiting the "suspicious feeding of stray animals" state by twice its original size. Images of the same size and position are then cropped from these frames using this ROI and arranged in chronological order to form a total of 6 seconds of offline video, which is saved locally for subsequent judgment by the action recognition module of the refined video stream.

[0066] Furthermore, in order to achieve more refined action recognition and detection, a C3D network structure can be used for the detection of temporal video streams, and the network and training process can be optimized and adapted.

[0067] During training, open-source datasets can be accessed to pre-train the artificial intelligence analysis model. More specifically, open-source datasets can be obtained by mixing real-world simulated data with generative data from large AI models.

[0068] To ensure sufficient diversity and realism in the data, the following methods were employed for data collection and processing: A certain amount of simulated real-world data was collected, taking into account factors such as the range of motion during feeding stray animals, camera angles, and ambient lighting. This data was then used as a model to generate generative video data corresponding to the actions performed by an AI analysis model. Through these steps, the generated data is no longer the previously mentioned 6-second offline video but a longer generated video, which is then randomly downsampled at a 3:5 ratio to create a 6-second randomly sampled video. Finally, reinforcement learning of the model is performed based on expert experience calibration.

[0069] Furthermore, after step S30 above, the method of this embodiment further includes: S300: Converts multiple time-series consecutive initial images into an RGBA four-channel image.

[0070] Specifically, the initial image is converted from the original RGB three-channel data to RGBA four-channel data. This means that in addition to the original visible light three-channel data, a human body mask image with a value range of 0-1 is added. More specifically, the human body bounding box in each frame of the video is used as a Region of Interest (ROI), and the corresponding ROI position in the A channel is set to 1, while other positions are set to 0. This method allows the deep neural network to develop attention capabilities during relearning, improving the network's accuracy.

[0071] S301: Input the RGBA four-channel image and the judgment result of the current frame into the artificial intelligence analysis model to obtain the comprehensive judgment result of multiple frames.

[0072] The artificial intelligence analysis model is configured as follows: First, the RGBA four-channel image is sampled to obtain a sampled image. Then, the sampled image and the judgment result of the current frame are convolved to obtain temporal image features. Finally, based on the temporal image features and a preset binary cross-entropy loss function, a multi-frame comprehensive judgment result is predicted.

[0073] For example, suppose the data format input to the artificial intelligence analysis model is 16 3 768 448, where the first column is the frame number, the second column is the number of image channels, the third column is the image width, and the fourth column is the image length. 16 frames are obtained by Gaussian sampling of a 6-second video.

[0074] To improve the inference speed of the network in the model, 2D convolution is used in the first half of the model network to obtain independent image features for each frame. By stacking 3 layers of convolution and 2 downsampling operations, the final result is a 16-bit image. 3 768 Feature data of 448.

[0075] Then, the independent feature data are passed through three layers of SparseConv3D convolution with downsampling to establish connections between the features of each frame, resulting in temporal image features with a data format of 16. 512 twenty four 14.

[0076] Finally, by performing pooling operations on the timing channels, we can obtain the format 512. twenty four The final feature of 14 is mapped to 2 through a fully connected FC (Fully Connected Network). The prediction node is 1, and then the BCE binary cross-entropy loss can guide the model to make more accurate predictions.

[0077] Conversely, the same binary cross-entropy loss can be used in the model training process.

[0078] More specifically, the binary cross-entropy loss function described above is:

[0079] in, This represents the actual label, which ranges from 0 to 1. This represents the probability predicted by the model, and it takes values ​​between 0 and 1. This represents the binary cross-entropy loss value.

[0080] Based on the above model prediction, if the multi-frame comprehensive judgment result is still to feed stray animals, then step S40 is executed.

[0081] S40: Based on the comprehensive judgment results of multiple frames, drive the inspection robot to perform the corresponding duty tasks.

[0082] Specifically, the robot is driven to issue voice prompts and risk warnings to the relevant individuals who feed the animals, and the information is recorded in the background. If the robot identifies the individual as a foreign tourist, it can switch the voice prompts to an internationally recognized language.

[0083] Based on the above technical description, the method of this embodiment has at least the following advantages: Firstly, the method in this embodiment uses robotic inspection to patrol public places such as residential areas and squares. The inspection robot, equipped with a common, inexpensive camera, can efficiently detect whether people are feeding stray animals within the inspection route. Compared to using fixed cameras, this method not only greatly expands the detection range but also increases deployment flexibility.

[0084] Secondly, the method of this embodiment has higher accuracy and efficiency in the detection process, and can be directly deployed on edge computing devices, with real-time detection capabilities.

[0085] Thirdly, the method in this embodiment specifically optimizes the human keypoint Alphapose framework for robot deployment platforms, selects keypoints at appropriate locations, and retrains Alphapose on a dataset formed by merging multiple datasets. The keypoint detection logic algorithm is not only fast but also extremely accurate, effectively supporting subsequent logical filtering. In addition, the method in this embodiment effectively improves the performance of the action recognition C3D network structure, reduces the overall inference speed of the network while basically maintaining the accuracy, and also provides great convenience for edge deployment.

[0086] Example 2 like Figure 3 As shown, this embodiment also discloses a feeding behavior detection system for stray animals based on an inspection robot, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the feeding behavior detection method for stray animals based on an inspection robot described in Embodiment 1 is implemented.

[0087] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions that can be stored or otherwise maintained by such a computer-readable medium.

[0088] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0089] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for detecting the feeding behavior of stray animals based on an inspection robot, characterized in that, include: Initial images are continuously collected along the predetermined inspection path; Based on a preset real-time target detection model, the initial image is analyzed to obtain features of ownerless animal aggregation and human behavior. The features of the unclaimed animals and the features of human behavior are input into a preset artificial intelligence analysis model to obtain a multi-frame comprehensive judgment result; Based on the comprehensive judgment results of the multiple frames, the inspection robot is driven to perform the corresponding duty tasks.

2. The method for detecting the feeding behavior of stray animals based on an inspection robot according to claim 1, characterized in that, Based on a preset real-time target detection model, the initial image is parsed, including: From the initial image, select several ownerless animal frames; The cluster centroid is calculated based on the center of several bounding boxes without an active action. Calculate the average Euclidean distance from the center of all unclaimed animal frames to the aggregate centroid; Based on the average Euclidean distance, the aggregation characteristics of the unclaimed animals are marked as aggregated or non-aggregated.

3. The method for detecting the feeding behavior of stray animals based on an inspection robot according to claim 2, characterized in that, Based on a preset real-time target detection model, the initial image is parsed, which further includes: From the initial image, pick the human detection box that is closest to the ownerless animal frame; Calibrate the key nodes and spatial coordinates of the human body detection frame; The key nodes are filtered based on the visual features of the initial image or the spatial coordinates to obtain multiple reliable key nodes; The spatial relationship between multiple trusted key nodes is calculated to obtain human behavioral characteristics that characterize the node spacing distance and trunk flexion and extension angle.

4. The method for detecting the feeding behavior of stray animals based on an inspection robot according to claim 3, characterized in that, The human behavioral characteristics include the wrist-hip distance; the unclaimed animal aggregation characteristics and the human behavioral characteristics are input into a preset artificial intelligence analysis model to obtain multi-frame comprehensive judgment results, including: If the animal aggregation characteristic is aggregation, and the trunk flexion-extension angle is greater than a preset first angle threshold, the following steps are performed: Determine whether the wrist-hip interval is greater than a preset second interval distance. If so, obtain the current frame determination result used to characterize the feeding state.

5. The method for detecting the feeding behavior of stray animals based on an inspection robot according to claim 4, characterized in that, After obtaining the current frame determination result, the method further includes: Convert multiple temporally consecutive initial images into RGBA four-channel images; The RGBA four-channel image and the current frame judgment result are input into the artificial intelligence analysis model to obtain the multi-frame comprehensive judgment result.

6. The method for detecting the feeding behavior of stray animals based on an inspection robot according to claim 5, characterized in that, The artificial intelligence analysis model is configured as follows: The RGBA four-channel image is sampled to obtain a sampled image; Convolve the sampled image and the current frame judgment result to obtain temporal image features; Based on the temporal image features and the preset binary cross-entropy loss function, a multi-frame comprehensive judgment result is predicted.

7. The method for detecting the feeding behavior of stray animals based on an inspection robot according to claim 1, characterized in that, The real-time object detection model adopts a YOLOv8-based detection model; the real-time object detection model uses the ReLU activation function; the real-time object detection model also replaces the original YOLOv8 upsample operator with a combination of operators including deconvolution operator, Batch Normalization operator, and ReLU activation operator.

8. The method for detecting the feeding behavior of stray animals based on an inspection robot according to claim 1, characterized in that, If the animal aggregation feature is aggregation, proceed with the human behavior feature detection step; otherwise, terminate the human behavior feature detection step for the current frame.

9. The method for detecting the feeding behavior of stray animals based on an inspection robot according to claim 1, characterized in that, Also includes: The artificial intelligence analysis model is pre-trained.

10. A feeding behavior detection system for stray animals based on an inspection robot, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the feeding behavior detection method for stray animals based on an inspection robot as described in any one of claims 1-9 is implemented.

Images