Abnormal garbage dumping traceability method and system based on computer vision stack index

Abstract

This invention provides a method and system for tracing abnormal waste disposal based on a computer vision stacked index, belonging to the field of abnormal waste disposal tracing technology. The method includes acquiring a target baseline image and full monitoring video of abnormal waste through a visual monitoring system, preprocessing the target baseline image to generate a target feature vector; identifying and extracting waste disposal events using a target detection model, arranging the waste disposal events in ascending order according to the waste disposal timestamp to generate an original sequence of waste disposal events; comparing and analyzing steady-state image frames using an inter-frame difference algorithm, filtering incremental object images and extracting incremental object feature vectors to construct a stacked temporal feature index library; traversing the stacked temporal feature index library in reverse order, performing feature matching through an early stopping mechanism, and generating target disposal tracing results. This enables accurate indexing of abnormal waste disposal events and rapid tracing of disposal personnel, improving the efficiency and accuracy of abnormal waste disposal tracing.

Description

Technical Field

[0001] This invention relates to the field of abnormal waste disposal tracing technology, and in particular to an abnormal waste disposal tracing method and system based on computer vision stack indexing. Background Technology

[0002] With the advancement of refined urban management, household waste sorting and recycling has become an important part of community management. In practice, waste collection personnel manually count or randomly inspect waste to identify those with incorrect sorting or abnormal weight during overflow bin handling or scheduled collection. Tracing the responsible party for disposing of such waste has become a key task in ensuring waste sorting.

[0003] However, existing methods for tracing abnormal waste disposal mainly rely on manual, linear review of surveillance footage from the collection point over the past few hours or even days, or frame-by-frame playback using simple motion detection. Since waste collection is a continuous accumulation process, and the surveillance videos span a wide range of durations, it is difficult for managers to directly determine the specific time of disposal for a particular piece of abnormal waste. This retrieval method is not only extremely labor-intensive but also highly inefficient. In high-frequency disposal scenarios, it is prone to missed detections or misjudgments due to visual fatigue. Furthermore, existing automated video analysis technologies often lack the utilization of physical scene patterns, and their commonly used full-data traversal retrieval strategy easily leads to significant waste of computational resources, failing to meet the needs of rapid on-site tracing.

[0004] Therefore, it is necessary to provide a method and system for tracing abnormal garbage disposal based on computer vision stack indexing to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for tracing the source of abnormal garbage disposal based on computer vision stacked indexing. This method solves the problems of existing abnormal garbage tracing methods relying on manual retrieval, which is inefficient, and the redundant computing resources of automated video analysis technology.

[0006] The present invention provides a method for tracing the source of abnormal garbage disposal based on computer vision stacked indexing, the method comprising: The system acquires target reference images and full monitoring videos of abnormal waste through a visual monitoring system, and preprocesses the target reference images to generate target feature vectors. A target detection model is used to identify and extract garbage disposal events from the full amount of monitoring video. The garbage disposal events are arranged in ascending order according to the garbage disposal timestamp to generate the original sequence of garbage disposal events. Based on the original sequence of the garbage disposal event, the steady-state image frames are compared and analyzed by the inter-frame difference algorithm, incremental object images are filtered and incremental object feature vectors are extracted, and a stacked temporal feature index library is constructed. The stacked temporal feature index library is traversed in reverse order, and the incremental object feature vector is matched with the target feature vector through an early stopping mechanism to generate target delivery tracing results.

[0007] Preferably, the step of acquiring the target reference image and full monitoring video of abnormal waste through the visual monitoring system, and preprocessing the target reference image to generate a target feature vector, specifically includes: The visual monitoring system acquires target reference images of abnormal waste from multiple angles in real time and uploads them to a smart terminal. The smart terminal automatically records the acquisition timestamp, abnormal waste collection point information, and acquisition angle information of the target reference image. Based on the collection timestamp, the smart terminal retrieves the full monitoring video through the SDK interface of the visual monitoring system according to the abnormal garbage collection point information; The target reference image is processed sequentially using the gray world algorithm for color balancing, bilinear interpolation for scaling, and Gaussian blur algorithm for noise suppression, resulting in a standard target reference image. The local texture features, global contour features, and color distribution features of the standard target reference image are extracted by the shallow, medium, and deep convolutional layers of the ResNet model, respectively. The local texture features, global contour features, and color distribution features are then subjected to numerical vector mapping and normalization using linear transformation and activation functions to generate the target feature vector.

[0008] Preferably, the step of using a target detection model to identify and extract garbage disposal events from the full amount of surveillance video specifically includes: The full monitoring video is input into the target detection model. The video frame decoding module based on the target detection model parses the full monitoring video frame by frame into a continuous image frame sequence. The bounding box regression algorithm is used to identify the garbage containment image frame sequence in the continuous image frame sequence, and the garbage containment image frame sequence is extracted by pixel mask counting technology. A preset image pixel fluctuation threshold and fluctuation duration T are defined. Based on the start and end timestamps of the image pixel fluctuations in the waste containment image frame sequence, if the rate of change of the image pixel fluctuations in the waste containment image frame sequence is lower than the image pixel fluctuation threshold and the image pixel fluctuation duration is not lower than the fluctuation duration T, then the waste containment image frame sequence is determined to be in a stable state. If the rate of change of image pixels in the waste containment image frame sequence is higher than the image pixel fluctuation threshold, then the waste containment image frame sequence is determined to be in a changed state, and the waste containment image frame sequence is marked as the waste disposal event; After determining that the waste containment image frame sequence is in a changed state, if the rate of change of the image pixel fluctuation of the waste containment image frame sequence is lower than the image pixel fluctuation threshold, and the duration of the image pixel fluctuation is not lower than the duration of the fluctuation T, then it is determined that the waste containment image frame sequence has completed the change and returned to a stable state.

[0009] Preferably, the step of arranging the waste disposal events in ascending order according to the waste disposal timestamp to generate the original sequence of waste disposal events specifically includes: The garbage disposal event and the corresponding image frame number range are uploaded to the smart terminal for structured encapsulation to obtain a structured garbage disposal event; The smart terminal extracts the start timestamp of the waste disposal event and converts it into a standard start time value. Using the standard start time value as the sorting basis, the structured waste disposal events are sorted in ascending order using a timestamp sorting algorithm to generate the original sequence of the waste disposal events.

[0010] Preferably, the step of comparing and analyzing steady-state image frames based on the original sequence of the garbage disposal event, filtering incremental object images, and extracting incremental object feature vectors specifically includes: Based on the image frame number corresponding to the garbage disposal event and the image frame rate of the full monitoring video The range of old steady-state image frame numbers is obtained as follows: The range of new steady-state image frame numbers is ; Traverse the range of old steady-state image frame numbers in reverse order The last frame whose image pixel fluctuation rate is lower than the image pixel fluctuation threshold is selected as the old steady-state image frame, and the number range of the new steady-state image frames is traversed in ascending order. The first frame whose image pixel distribution volatility is lower than the image pixel volatility threshold is selected as the new steady-state image frame; A preset pixel difference threshold is set, and the old steady-state image frame and the new steady-state image frame are compared at the pixel level using the inter-frame difference algorithm to calculate the pixel difference. If the pixel difference is greater than or equal to the pixel difference threshold, the pixel corresponding to the pixel difference in the new steady-state image frame is retained to generate a difference image. The difference image is then subjected to binarization and morphological filtering in sequence, and the incremental object image is filtered using a connected component filtering algorithm. The incremental object image is input into a convolutional neural network model for feature encoding. A sliding filter is used to capture the local visual features of the incremental object image to generate a multi-channel feature map. A pooling layer is used to compress the size of the multi-channel feature map to generate a high-channel feature map. A fully connected layer is used to map the high-dimensional feature map into a fixed-dimensional feature vector of the incremental object.

[0011] Preferably, the construction of the stacked time-series feature index library specifically includes: Each incremental object feature vector is data-bound using metadata tags to form an index data item. The index data item includes the incremental object feature vector, the start timestamp of the garbage disposal event corresponding to the incremental object feature vector, and the image frame index of the garbage disposal event corresponding to the incremental object feature vector in the full-volume monitoring video. Using a key-value pair storage mode, the index data items are pushed into the stacked time-series feature index library in ascending order according to a preset stacked storage rule.

[0012] Preferably, the step of reverse traversing the stacked temporal feature index library and performing feature matching between the incremental object feature vector and the target feature vector through an early stopping mechanism specifically includes: Based on the reverse depth-first search strategy, the incremental object feature vector U is read one by one from the top of the stacked temporal feature index library, and the cosine similarity between the incremental object feature vector U and the target feature vector V is calculated. The corresponding calculation formula is as follows: In the formula, This represents the dimension of the incremental object feature vector, and the dimension of the target feature vector is equal to that of the incremental object feature vector; This represents the i-th element of the incremental object feature vector; This represents the i-th element of the target feature vector; A pre-set confidence threshold is used to adjust the cosine similarity through the early stopping mechanism. Compare with the confidence threshold, if the cosine similarity If the confidence level is greater than or equal to the confidence threshold, the feature matching is determined to be successful. The start timestamp of the garbage disposal event corresponding to the incremental object feature vector is locked as the target disposal timestamp, and the image frame index of the garbage disposal event corresponding to the incremental object feature vector in the full monitoring video is locked as the target monitoring video index. Feature matching is then stopped.

[0013] Preferably, the target deployment tracing result is generated based on the target deployment timestamp and the target surveillance video index, specifically including: Based on the target delivery timestamp and the target monitoring video index, locate the corresponding full monitoring video, and use the target delivery timestamp as the cutoff timestamp to backtrack the full monitoring video according to a preset behavior cycle to generate a backtracking monitoring video; Human body recognition algorithms are used to identify human objects in the retrospective surveillance video, and facial feature images and body posture feature images of the human objects are extracted. Based on the target deployment timestamp, the target baseline image, the retrospective monitoring video, and the facial and body feature images of the human object are aggregated and associated to generate the target deployment tracing result.

[0014] An abnormal garbage disposal tracing system based on computer vision stack indexing, the system comprising: The image acquisition and preprocessing module is used to acquire target reference images and full monitoring videos of abnormal garbage through the visual monitoring system, and to preprocess the target reference images to generate target feature vectors. The waste disposal identification module is used to identify and extract waste disposal events from the full monitoring video using a target detection model, arrange the waste disposal events in ascending order according to the waste disposal timestamp, and generate the original sequence of waste disposal events. The feature index construction module is used to compare and analyze steady-state image frames based on the original sequence of the garbage disposal event, filter incremental object images and extract incremental object feature vectors, and construct a stacked time-series feature index library. The target delivery tracing module is used to traverse the stacked time-series feature index library in reverse order, and perform feature matching between the incremental object feature vector and the target feature vector through an early stopping mechanism to generate target delivery tracing results.

[0015] Compared with existing technologies, the abnormal garbage disposal tracing method and system based on computer vision stack indexing provided by this invention has the following beneficial effects: This invention acquires target baseline images and full-scale monitoring videos of abnormal waste through a visual monitoring system. The target baseline images are preprocessed to generate target feature vectors. A target detection model is used to identify and extract waste disposal events from the full-scale monitoring videos. These events are arranged in ascending order according to their timestamps to generate an original sequence of waste disposal events. Based on this original sequence, a frame-difference algorithm is used to compare and analyze steady-state image frames, filter incremental object images, and extract incremental object feature vectors to construct a stacked temporal feature index library. The stacked temporal feature index library is then traversed in reverse order, and an early stopping mechanism is used to match the incremental object feature vectors with the target feature vectors to generate target disposal tracing results. This enables accurate indexing of abnormal waste disposal events and rapid tracing of disposal personnel, improving the efficiency and accuracy of abnormal waste disposal tracing.

[0016] This invention identifies and extracts garbage disposal events from a full suite of surveillance videos using a target detection model. It arranges these events in ascending order of timestamps, generating a raw sequence of events and precisely filtering out meaningless still images to reduce data redundancy. Furthermore, by utilizing a stacked temporal feature index library and leveraging the physical principle of garbage accumulation (last-in, first-out), it binds incremental object feature vectors to the disposal time dimension. This addresses the shortcomings of existing automated analysis techniques, such as neglecting scene physics and inefficient retrieval strategies, enabling efficient storage and rapid retrieval of feature data. Finally, this invention employs reverse depth... The priority detection strategy and early stop mechanism prioritize matching recent high-probability waste disposal events, avoiding the waste of computing resources caused by traversing the entire monitoring video frame by frame. This solves the problems of time-consuming, missed, and misjudged searches caused by the traditional needle-in-a-haystack approach to tracing abnormal waste disposal, significantly improving the response speed of abnormal waste disposal tracing. This invention automates the entire process from image acquisition, event extraction, stacked indexing to disposal tracing, without the need for manual screening, greatly reducing labor costs. At the same time, through precise comparison of incremental object feature vectors and target feature vectors, the reliability of tracing results is improved, which is conducive to the refined management of urban waste classification. Attached Figure Description

[0017] Figure 1 A flowchart of an abnormal garbage disposal tracing method based on computer vision stacked index provided in an embodiment of the present invention; Figure 2 This is a system block diagram of an abnormal waste disposal tracing system based on computer vision stacked indexing provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 embodiments of the present invention, and not all embodiments. 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.

[0019] like Figure 1 The diagram shown is a flowchart of an abnormal garbage disposal tracing method based on computer vision stacked indexing provided in an embodiment of the present invention. Figure 1The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. Steps S1 to S4 are detailed as follows: S1, acquire the target reference image and full monitoring video of the abnormal garbage through the visual monitoring system, and preprocess the target reference image to generate the target feature vector; Abnormal waste refers to household waste that exhibits problems such as incorrect sorting or abnormal weight during overflow disposal or scheduled collection, failing to meet waste sorting standards or possessing abnormal attributes. The target baseline image refers to a photograph of the abnormal waste taken on-site by a visual monitoring system when it is discovered. Full-volume monitoring video refers to continuous monitoring video data retrieved from the visual monitoring system of the waste collection point to which the abnormal waste belongs, based on the acquisition timestamp of the target baseline image. This data covers potential disposal times of the abnormal waste, ensuring no abnormal waste disposal is missed. The target feature vector refers to the extracted visual features of the target baseline image, such as local texture, global contour, and color distribution, used to accurately correlate abnormal waste with disposal events.

[0020] Understandably, retrieving all monitoring videos based on the collection timestamp accurately pinpoints the possible time range of abnormal waste disposal, avoiding redundancy caused by irrelevant video data. The target reference image captured on-site is directly associated with the abnormal waste to be traced. Through preprocessing, not only can interference factors such as shooting angle, lighting, and noise be eliminated, but the concrete target reference image can also be transformed into a computer-computable target feature vector, achieving a precise mapping from abnormal waste in the physical world to feature vectors in the digital world.

[0021] The process of acquiring a target reference image and full monitoring video of abnormal waste through a visual monitoring system, and preprocessing the target reference image to generate a target feature vector, specifically includes: The visual monitoring system acquires target reference images of abnormal waste from multiple angles in real time and uploads them to a smart terminal. The smart terminal automatically records the acquisition timestamp, abnormal waste collection point information, and acquisition angle information of the target reference image. Based on the collection timestamp, the smart terminal retrieves the full monitoring video through the SDK interface of the visual monitoring system according to the abnormal garbage collection point information; The target reference image is processed sequentially using the gray world algorithm for color balancing, bilinear interpolation for scaling, and Gaussian blur algorithm for noise suppression, resulting in a standard target reference image. The local texture features, global contour features, and color distribution features of the standard target reference image are extracted by the shallow, medium, and deep convolutional layers of the ResNet model, respectively. The local texture features, global contour features, and color distribution features are then subjected to numerical vector mapping and normalization using linear transformation and activation functions to generate the target feature vector.

[0022] First, a target reference image captured from a single angle cannot fully represent the visual characteristics of abnormal waste. Therefore, a visual monitoring system is used to acquire target reference images of abnormal waste in real time from multiple angles, including the front, side, and top, and upload them to a smart terminal immediately. The smart terminal simultaneously records the acquisition timestamp, acquisition angle information, and abnormal waste collection point information, such as the abnormal waste collection point number and its specific location coordinates, thus achieving structured storage of the target reference images.

[0023] To ensure complete coverage of potential abnormal waste disposal periods, based on the collection timestamp, the smart terminal retrieves the full range of continuous monitoring video from the abnormal waste collection point from T1 hours before the collection timestamp to T2 hours after the collection timestamp through the SDK interface of the visual monitoring system, while ensuring the stability and compatibility of data transmission, so as to ensure that no possible disposal scenes are missed.

[0024] It should be noted that T1 and T2 can be set according to the actual scenario, such as T1=12 hours and T2=1 hour, which can cover the most likely time period for delivery and avoid excessive redundancy of video data.

[0025] Subsequently, the target reference image undergoes a three-step preprocessing process. First, a grayscale world algorithm is used for color balancing. Using the average grayscale value of the target reference image as a fixed value, the pixel values ​​of each channel are adjusted to achieve color uniformity, effectively eliminating color distortion caused by illumination. Second, bilinear interpolation is used for scaling, uniformly scaling the target reference image to a preset size, balancing feature extraction efficiency and detail preservation, ensuring consistency with the image frame size of the entire surveillance video. Finally, a Gaussian blur algorithm is used for noise suppression. By setting an appropriate Gaussian kernel size, noise reduction is achieved while avoiding blurring key features, filtering out random noise in the target reference image, and outputting a clear, standardized target reference image.

[0026] Furthermore, feature preprocessing is performed on the standard target benchmark image. The ResNet model possesses strong feature extraction and gradient propagation capabilities, enabling it to accurately capture the visual features of abnormal waste. Shallow convolutional layers capture local texture features of the standard target benchmark image, such as the texture, wrinkles, and material details of the abnormal waste packaging. Mid-level convolutional layers extract global contour features of the standard target benchmark image, such as the overall shape, size ratio, and contour edges of the abnormal waste. Deep convolutional layers mine color distribution features of the standard target benchmark image, such as the main color, color partitions, and gradient patterns of the abnormal waste. To transform multidimensional visual features into a computer-computable numerical form, linear transformations are used to unify the dimensions and numerically map local texture features, global contour features, and color distribution features. Then, activation functions are used to enhance the non-linear expressive power of the features and normalization is performed, mapping the feature vector values ​​to the [0,1] interval. Finally, a target feature vector with fixed dimensions that balances feature expressive power and computational efficiency is generated.

[0027] The above method transforms the target baseline image of abnormal waste in physical form into a standardized target feature vector, and solves the problem of difficulty in associating physical objects with videos in traditional source tracing.

[0028] S2, Use a target detection model to identify and extract garbage disposal events from the full monitoring video, arrange the garbage disposal events in ascending order according to the garbage disposal timestamp, and generate the original sequence of garbage disposal events; Among them, a garbage disposal event refers to the complete sequence of image frames corresponding to the process in which the image pixel distribution of the garbage-collecting area fluctuates significantly and eventually stabilizes in the full monitoring video. This sequence is used to fragment and mark a garbage disposal behavior, eliminating meaningless still images. The original sequence of garbage disposal events refers to the structured set of garbage disposal events formed by arranging all identified garbage disposal events in ascending order of their timestamps, containing the timestamp range and image frame range information for each garbage disposal event.

[0029] It should be noted that full-volume surveillance video contains a large amount of static footage showing no change in waste disposal areas. Direct analysis of this footage would result in a significant waste of computational resources and would fail to accurately pinpoint the specific time period of waste disposal. By using a target detection model to filter out waste disposal events containing only waste disposal behavior from continuous full-volume surveillance video and arranging them in chronological order, the lengthy and unstructured full-volume surveillance video can be automatically transformed into a structured, chronologically ordered original sequence of waste disposal events. This effectively solves the problems of low efficiency and missed detections / false detections associated with traditional manual playback of full-volume surveillance video, achieving dimensionality reduction of full-volume surveillance video and refined, intelligent management of waste disposal behavior.

[0030] The process of identifying and extracting garbage disposal events from the full volume of surveillance video using a target detection model specifically includes: The full monitoring video is input into the target detection model. The video frame decoding module based on the target detection model parses the full monitoring video frame by frame into a continuous image frame sequence. The bounding box regression algorithm is used to identify the garbage containment image frame sequence in the continuous image frame sequence, and the garbage containment image frame sequence is extracted by pixel mask counting technology. A preset image pixel fluctuation threshold and fluctuation duration T are defined. Based on the start and end timestamps of the image pixel fluctuations in the waste containment image frame sequence, if the rate of change of the image pixel fluctuations in the waste containment image frame sequence is lower than the image pixel fluctuation threshold and the image pixel fluctuation duration is not lower than the fluctuation duration T, then the waste containment image frame sequence is determined to be in a stable state. If the rate of change of image pixels in the waste containment image frame sequence is higher than the image pixel fluctuation threshold, then the waste containment image frame sequence is determined to be in a changed state, and the waste containment image frame sequence is marked as the waste disposal event; After determining that the waste containment image frame sequence is in a changed state, if the rate of change of the image pixel fluctuation of the waste containment image frame sequence is lower than the image pixel fluctuation threshold, and the duration of the image pixel fluctuation is not lower than the duration of the fluctuation T, then it is determined that the waste containment image frame sequence has completed the change and returned to a stable state.

[0031] Among them, the garbage containment image frame sequence refers to the set of image frames that contain garbage containment areas, such as garbage cans and garbage disposal openings, located and filtered in a continuous image frame sequence, which are used to eliminate background interference and accurately capture garbage disposal events.

[0032] It should be noted that the garbage disposal incidents only occur within the garbage collection area. Extracting the garbage collection image frame sequence is a prerequisite for accurately identifying garbage disposal incidents. If the image pixel fluctuation analysis is performed directly on the continuous image frame sequence of the entire monitoring video, irrelevant changes in the background will lead to a large number of misjudgments, such as pedestrians passing by or changes in lighting. Identifying and extracting the garbage collection image frame sequence can improve the accuracy and efficiency of garbage disposal incident identification.

[0033] First, by using the video frame decoding module of the target detection model, the continuous full monitoring video is parsed frame by frame into a discrete continuous image frame sequence that can be analyzed frame by frame by the computer, based on the image frame rate of the full monitoring video. This enables precise frame-by-frame analysis of the full monitoring video, avoiding the problem that continuous full monitoring video is difficult to split and cannot locate garbage disposal events, and ensuring that no potential disposal behavior is missed.

[0034] Secondly, leveraging the frame-by-frame independence of a continuous image frame sequence, the waste containment area is accurately located, avoiding interference from the background. Using a bounding box regression algorithm, based on the visual features of the waste containment area learned during the object detection model training phase, such as contour, size, and positional patterns, the bounding box of the waste containment area is accurately fitted to each image frame in the continuous image frame sequence, thus locking the spatial coordinates of the waste containment area. Then, using pixel mask counting technology, the image pixels within the bounding box of the waste containment area are marked and counted, and image frames containing the complete waste containment area are selected to form a waste containment image frame sequence.

[0035] Furthermore, to distinguish between a stable state with no litter disposal and a changing state with litter disposal, scientific judgment rules need to be established. The image pixel fluctuation threshold should be dynamically adjusted based on the camera resolution in the visual monitoring system and the environmental complexity of the litter collection area. Presetting the duration of fluctuations ensures the accuracy of stable state determination; for example, a brief swaying of litter caused by wind will not be misjudged as a litter disposal event. By calculating the difference between the sum of the current image pixel values ​​of the litter collection image frame and the sum of the pixel values ​​of the previous image frame, and then dividing by the sum of the pixel values ​​of the previous image frame, the image pixel fluctuation rate of the litter collection image frame sequence can be calculated, which can intuitively reflect changes in the quantity or shape of litter within the litter collection area.

[0036] Specifically, when the rate of change of image pixels in the waste containment image frame sequence is lower than the image pixel fluctuation threshold and remains in this fluctuation state for more than a preset fluctuation duration, the waste containment image frame sequence is determined to be in a stable state. When the rate of change of image pixels in any waste containment image frame sequence is higher than the image pixel fluctuation threshold, the waste containment image frame sequence is determined to be in a changed state, and this segment of the frame sequence is marked as a waste disposal event. When the rate of change of image pixels in the changed waste containment image frame sequence returns to below the image pixel fluctuation threshold and remains in this fluctuation state for more than a preset fluctuation duration, the waste disposal is determined to be completed, the waste image area returns to a stable state, and the identification of a complete waste disposal event is completed.

[0037] The step of arranging the waste disposal events in ascending order according to the waste disposal timestamp to generate the original sequence of waste disposal events specifically includes: The garbage disposal event and the corresponding image frame number range are uploaded to the smart terminal for structured encapsulation to obtain a structured garbage disposal event; The smart terminal extracts the start timestamp of the waste disposal event and converts it into a standard start time value. Using the standard start time value as the sorting basis, the structured waste disposal events are sorted in ascending order using a timestamp sorting algorithm to generate the original sequence of the waste disposal events.

[0038] Among them, structured waste disposal events can ensure the readability, compatibility, and quick access of waste disposal event data.

[0039] Understandably, converting the start timestamps of waste disposal events to standard start timestamps in milliseconds can eliminate the impact of time format differences on the ascending order sorting of structured waste disposal events. Using a timestamp ascending order sorting algorithm can efficiently process complex structured waste disposal events, ensuring that the original sequence of waste disposal events is accurately arranged from earliest to latest according to the time of disposal.

[0040] In addition, simultaneously encapsulating the image frame number range corresponding to the garbage disposal event helps to locate the steady-state image frame number range, avoids retracing the entire monitoring video to find the corresponding frame interval, and improves the continuity of the abnormal garbage disposal tracing process.

[0041] Overall, it not only achieves accurate identification and effective separation of waste disposal events, eliminating invalid still images from the full range of surveillance videos and solving the problem of wasted computing resources caused by full analysis in traditional automated video analysis technology, but also generates a clear and complete original sequence of waste disposal events through structured encapsulation and ascending order, significantly reducing the cost of manual intervention and realizing refined management of waste disposal behavior.

[0042] S3, based on the original sequence of the garbage disposal event, the steady-state image frames are compared and analyzed by the inter-frame difference algorithm, the incremental object images are filtered and the incremental object feature vectors are extracted, and a stacked time-series feature index library is constructed. Among them, a steady-state image frame refers to an image frame in the sequence of waste containment image frames that is in a stable state. It includes old steady-state image frames and new steady-state image frames. An old steady-state image frame is the last image frame in which the waste containment area is in a stable state before the waste disposal event occurs. A new steady-state image frame is the first image frame in which the waste containment area returns to a stable state after the waste disposal event ends.

[0043] Incremental object images refer to images containing only newly added waste from the current waste disposal event, selected by comparing old and new steady-state image frames. Incremental object feature vectors are standardized numerical vectors obtained after feature encoding of incremental object images, used for similarity matching with target feature vectors. A stacked temporal feature index library is an index set formed by binding incremental object feature vectors to corresponding waste disposal event data using a stacked storage structure, adapting to the physical accumulation patterns of waste.

[0044] Understandably, the original sequence of garbage disposal events only locks in the time period and image frame range of the events. However, the image frame sequence corresponding to each event includes both newly added garbage and previously existing garbage. Directly extracting features would lead to feature confusion and prevent accurate association with the current disposal behavior. Constructing a data structure adapted to the physical accumulation patterns of garbage can achieve efficient retrieval. Therefore, by accurately locating steady-state image frames before and after the garbage disposal event, separating the incremental object images from the current disposal and extracting the incremental object feature vectors, and then constructing a stacked temporal feature index library, precise binding of single disposal, incremental garbage, feature vectors, and stacked indexes can be achieved.

[0045] The process of analyzing steady-state image frames based on the original sequence of the garbage disposal event using an inter-frame difference algorithm, filtering incremental object images, and extracting incremental object feature vectors specifically includes: Based on the image frame number corresponding to the garbage disposal event and the image frame rate of the full monitoring video The range of old steady-state image frame numbers is obtained as follows: The range of new steady-state image frame numbers is ; Traverse the range of old steady-state image frame numbers in reverse order The last frame whose image pixel fluctuation rate is lower than the image pixel fluctuation threshold is selected as the old steady-state image frame, and the number range of the new steady-state image frames is traversed in ascending order. The first frame whose image pixel distribution volatility is lower than the image pixel volatility threshold is selected as the new steady-state image frame; A preset pixel difference threshold is set, and the old steady-state image frame and the new steady-state image frame are compared at the pixel level using the inter-frame difference algorithm to calculate the pixel difference. If the pixel difference is greater than or equal to the pixel difference threshold, the pixel corresponding to the pixel difference in the new steady-state image frame is retained to generate a difference image. The difference image is then subjected to binarization and morphological filtering in sequence, and the incremental object image is filtered using a connected component filtering algorithm. The incremental object image is input into a convolutional neural network model for feature encoding. A sliding filter is used to capture the local visual features of the incremental object image to generate a multi-channel feature map. A pooling layer is used to compress the size of the multi-channel feature map to generate a high-channel feature map. A fully connected layer is used to map the high-dimensional feature map into a fixed-dimensional feature vector of the incremental object.

[0046] in, This refers to the image frame number at the start of the garbage disposal event. This refers to the image frame number at the end of a waste disposal event; both precisely define the range of image frames for a single waste disposal event.

[0047] First, based on the image frame number corresponding to the garbage disposal event and the image frame rate of the full monitoring video, the potential steady-state image frame range before and after the garbage disposal event is locked. For example, if the image frame rate of the full monitoring video is 30 frames / second, the preset fluctuation duration is 5 seconds, and the image frame number at the beginning of the garbage disposal event is 1000 frames, then the old steady-state image frame number range is [1000−5×30,1000−1]=[850,999] frames, ensuring that the image frames in the 5 seconds before the start of the garbage disposal event are covered. Similarly, the new steady-state range covers the image frames in the 5 seconds after the end of the garbage disposal event.

[0048] Secondly, the range of old steady-state image frame numbers is traversed in reverse order, that is, from... Towards The algorithm iterates through each frame, calculating the rate of change of image pixel fluctuations. The last frame whose rate of change of image pixel fluctuations first falls below a certain threshold, and which subsequently satisfies this condition, is selected as the old steady-state image frame. This old steady-state image frame represents the final stable state of the waste-accepting area before the waste disposal event, avoiding minor disturbances from earlier image frames. The algorithm then iterates through the range of new steady-state image frame numbers in ascending order, starting from... Towards The process iterates through the images, calculating the rate of change of image pixel fluctuations frame by frame. The first frame whose rate of change of image pixel fluctuations first falls below the image pixel fluctuation threshold and which is the first frame to meet this condition is selected as the new steady-state image frame. This new steady-state image frame represents the initial stable state of the waste containment area after the waste disposal event, reflecting the final result after the waste disposal.

[0049] Then, a preset pixel difference threshold is set, and the old steady-state image frame and the new steady-state image frame are compared pixel by pixel using the inter-frame difference algorithm. The pixel difference is calculated to be equal to the pixel value of the new steady-state image frame minus the pixel value of the old steady-state image frame. The larger the pixel difference, the more significant the change in image pixels, which is more likely to be newly added garbage. Pixels with pixel differences greater than or equal to the pixel difference threshold are retained to generate a difference image containing only the newly added area.

[0050] Furthermore, the difference image is optimized by first performing binarization, setting pixels with values ​​greater than 0 to 255 and the rest to 0 to highlight the contours of the difference image. Then, morphological operations of erosion and dilation are used to eliminate small noise points and connect broken contours. Finally, a connected component filtering algorithm is used to filter out interference regions by setting a minimum area threshold, resulting in a clean incremental object image.

[0051] Finally, the incremental object image is input into the convolutional neural network model, and local visual features are captured by the sliding filter, such as the addition of garbage texture and edge details, to generate a multi-channel feature map. The size of the multi-channel feature map is compressed by the max pooling layer and key features are retained to generate a high-channel feature map. Then, the high-dimensional feature map is mapped to a fixed-dimensional incremental object feature vector through a fully connected layer to ensure that the incremental object feature vector is directly compared with the target feature vector.

[0052] The construction of the stacked time-series feature index library specifically includes: Each incremental object feature vector is data-bound using metadata tags to form an index data item. The index data item includes the incremental object feature vector, the start timestamp of the garbage disposal event corresponding to the incremental object feature vector, and the image frame index of the garbage disposal event corresponding to the incremental object feature vector in the full-volume monitoring video. Using a key-value pair storage mode, the index data items are pushed into the stacked time-series feature index library in ascending order according to a preset stacked storage rule.

[0053] Metadata tags are the association identifiers assigned to incremental object feature vectors, used to accurately bind the incremental object feature vectors with the metadata of the corresponding waste disposal events, forming an inseparable associated whole. Index data items are standardized data units that store three-dimensional information of features, time, and location. They are the basic storage units of the stacked time-series feature index library, used to achieve the association and retrieval of incremental object feature vectors with key information of waste disposal events and the full monitoring video. Image frame indexes are used to quickly locate the specific image segment of the disposal event in the full monitoring video, avoiding the need to re-traverse the entire full monitoring video during retrieval.

[0054] It should be noted that the stacked time-series feature index library is constructed to adapt to the physical accumulation patterns of waste and ensure the efficiency of reverse-order priority retrieval. The binding of metadata tags enables the traceability of incremental object feature vectors and the correlation of waste disposal event information. The key-value pair storage mode ensures the efficiency of data retrieval, and the stacked storage rules directly determine the priority order during retrieval. The three elements work together to ensure the information integrity and retrieval efficiency of the stacked time-series feature index library.

[0055] Specifically, when performing metadata tag binding, a unique garbage disposal event association ID is assigned to each incremental object feature vector. This garbage disposal event association ID is consistent with the ID of the corresponding garbage disposal event in the original sequence of garbage disposal events. Based on the garbage disposal event association ID, the incremental object feature vector is bound to the start timestamp of the corresponding garbage disposal event and the image frame index in the corresponding full monitoring video, forming a complete index data item containing the incremental object feature vector, start timestamp, and image frame index, ensuring that each incremental object feature vector can be accurately associated with the corresponding garbage disposal event and the corresponding full monitoring video segment.

[0056] Subsequently, a key-value pair storage model is adopted for encapsulation. The garbage disposal event association ID in the index data item is used as the key, and the complete index data item containing the incremental object feature vector, start timestamp, and image frame index is used as the value. Each index data item is encapsulated as a key-value pair. This key-value pair storage model allows for quick retrieval of the corresponding value via the key, significantly shortening data retrieval time and avoiding a full traversal of all index data items. All index data items are pushed into the stack-based time-series feature index library in ascending order according to the stack-based storage rule of Last-In-First-Out (LIFO). All encapsulated index data items are pushed into the stack-based time-series feature index library in ascending order of the original sequence of garbage disposal events, i.e., in the order of the garbage disposal events occurring from earliest to latest. For example, the index data item corresponding to the earliest garbage disposal event is pushed to the bottom of the stack-based time-series feature index library, and the index data item corresponding to the latest garbage disposal event is pushed to the top. This ensures that when retrieving the stack-based time-series feature index library, the feature data of the incremental object feature vector corresponding to the most recently occurring garbage disposal event is prioritized, adapting to the actual scenario where abnormal garbage is likely to have been recently disposed of.

[0057] S4. Traverse the stacked temporal feature index library in reverse order, and perform feature matching between the incremental object feature vector and the target feature vector through the early stopping mechanism to generate target delivery tracing results.

[0058] The early stop mechanism refers to an optimization mechanism that immediately terminates subsequent searches once a matching result that meets the criteria is found during feature matching, in order to avoid invalid full traversal. The target waste disposal traceability result refers to a complete chain of evidence summarizing the physical information of abnormal waste, disposal time, video footage of the disposal process, and characteristics of the personnel who disposed of the waste, used to clarify the responsibility for the disposal of abnormal waste.

[0059] It should be noted that the stack-based time-series feature index library's last-in-first-out structure is highly compatible with the physical stacking pattern of garbage being added after it is disposed of. The reverse traversal combined with the early stopping mechanism can minimize the number of searches and solve the problems of low efficiency and wasted computing resources in traditional full traversal.

[0060] The reverse traversal of the stack-based temporal feature index library, and the feature matching of the incremental object feature vector with the target feature vector through an early stopping mechanism, specifically includes: Based on the reverse depth-first search strategy, the incremental object feature vector U is read one by one from the top of the stacked temporal feature index library, and the cosine similarity between the incremental object feature vector U and the target feature vector V is calculated. The corresponding calculation formula is as follows: In the formula, This represents the dimension of the incremental object feature vector, and the dimension of the target feature vector is equal to that of the incremental object feature vector; This represents the i-th element of the incremental object feature vector; This represents the i-th element of the target feature vector; A pre-set confidence threshold is used to adjust the cosine similarity through the early stopping mechanism. Compare with the confidence threshold, if the cosine similarity If the confidence level is greater than or equal to the confidence threshold, the feature matching is determined to be successful. The start timestamp of the garbage disposal event corresponding to the incremental object feature vector is locked as the target disposal timestamp, and the image frame index of the garbage disposal event corresponding to the incremental object feature vector in the full monitoring video is locked as the target monitoring video index. Feature matching is then stopped.

[0061] Specifically, the reverse depth-first detection strategy prioritizes matching from newest to oldest. Starting from the top of the stacked time-series feature index library, i.e., from the index data item corresponding to the latest garbage disposal event, incremental object feature vectors are read down one by one. The feature similarity is measured by calculating the cosine similarity between the incremental object feature vector and the target feature vector. The closer the cosine similarity is to 1, the more similar the features are. This accurately captures the visual feature association between abnormal garbage and incremental objects, avoiding misjudgments caused by differences in garbage size and shooting distance, thus adapting to the actual scenario of abnormal garbage being recently disposed of.

[0062] Subsequently, the preset confidence threshold is adjusted according to the accuracy requirements of the actual application. A higher confidence threshold results in higher matching accuracy, while a lower confidence threshold results in faster retrieval speed. This threshold is used to define the similarity of feature vectors. The cosine similarity is compared with the preset confidence threshold. If the cosine similarity is greater than or equal to the confidence threshold, it indicates that the newly added garbage corresponding to the incremental object feature vector is highly consistent with the abnormal garbage, and the match is determined to be successful. The start timestamp of the garbage disposal event in the index data item is immediately locked as the target disposal timestamp, and the image frame index of the full monitoring video in the index data item is locked as the target monitoring video index. All subsequent retrieval operations are terminated according to the early stop mechanism to avoid invalid calculations. If no match is found after traversing to the bottom of the stacked time-series feature index library, a retrieval failure is output, prompting manual intervention for verification.

[0063] Based on the target deployment timestamp and the target surveillance video index, the target deployment tracing result is generated, specifically including: Based on the target delivery timestamp and the target monitoring video index, locate the corresponding full monitoring video, and use the target delivery timestamp as the cutoff timestamp to backtrack the full monitoring video according to a preset behavior cycle to generate a backtracking monitoring video; Human body recognition algorithms are used to identify human objects in the retrospective surveillance video, and facial feature images and body posture feature images of the human objects are extracted. Based on the target deployment timestamp, the target baseline image, the retrospective monitoring video, and the facial and body feature images of the human object are aggregated and associated to generate the target deployment tracing result.

[0064] The preset behavior cycle refers to the duration of a complete waste disposal event, which is used to ensure that the retrospective monitoring video can fully capture the waste disposal event.

[0065] Understandably, by using the target disposal timestamp and the target monitoring video index, the monitoring video corresponding to the waste disposal event can be quickly located. Then, by tracing back the complete disposal process and extracting human characteristics, the abnormal waste, the abnormal waste disposal process, and the person who disposed of the abnormal waste can be linked to form a complete chain of evidence.

[0066] In practical applications, based on the target monitoring video index, the corresponding segment is directly extracted from the full monitoring video. Then, using the target disposal timestamp as the cutoff time, the footage before the abnormal waste disposal is supplemented by retrospective analysis, ensuring that the retrospective monitoring video can fully present the entire process of a person approaching and disposing of waste. Facial feature images include clear frontal or side views of the human body, and body posture images include the full-body posture of the human body, thus enabling the identification of the person disposing of waste. During the aggregation and association process, using the target disposal timestamp as an index, all target baseline images, retrospective monitoring videos, and facial and body posture images of the human object are presented in a structured manner, generating standardized target disposal tracing results.

[0067] Overall, by combining reverse traversal of the stack-based temporal feature index library, cosine similarity comparison, and early stopping mechanism, the system achieves rapid and accurate matching between abnormal waste and waste disposal events, solving the problem of inefficient retrieval in traditional abnormal waste disposal tracing. Furthermore, by backtracking monitoring videos, extracting human features, and associating information, the system generates complete target disposal tracing results, achieving fully automated tracing from the discovery of abnormal waste to the identification of abnormal waste disposal events and personnel, significantly reducing labor costs and providing technical support for the refined management of urban waste classification.

[0068] like Figure 2 The diagram shown is a system block diagram of an abnormal waste disposal tracing system based on computer vision stacked indexing provided in an embodiment of the present invention. The system includes: The image acquisition and preprocessing module is used to acquire target reference images and full monitoring videos of abnormal garbage through the visual monitoring system, and to preprocess the target reference images to generate target feature vectors. The waste disposal identification module is used to identify and extract waste disposal events from the full monitoring video using a target detection model, arrange the waste disposal events in ascending order according to the waste disposal timestamp, and generate the original sequence of waste disposal events. The feature index construction module is used to compare and analyze steady-state image frames based on the original sequence of the garbage disposal event, filter incremental object images and extract incremental object feature vectors, and construct a stacked time-series feature index library. The target delivery tracing module is used to traverse the stacked time-series feature index library in reverse order, and perform feature matching between the incremental object feature vector and the target feature vector through an early stopping mechanism to generate target delivery tracing results.

[0069] Figure 2 The system of the illustrated embodiment can be used to perform corresponding operations. Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0070] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the steps of the abnormal garbage disposal tracing method based on computer vision stack index as described above.

[0071] like Figure 3 The diagram shown is a hardware structure schematic of an electronic device according to an embodiment of the present invention. The electronic device 30 includes: a processor 31, a memory 32, and a computer program; wherein... The memory 32 is used to store the computer program, and the memory may also be flash memory. The computer program is, for example, an application program or functional module that implements the above method.

[0072] Processor 31 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0073] Alternatively, the memory 32 can be either standalone or integrated with the processor 31.

[0074] When the memory 32 is a device independent of the processor 31, the device may further include: Bus 33 is used to connect the memory 32 and the processor 31.

[0075] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of the abnormal garbage disposal tracing method based on computer vision stack indexing as described above.

[0076] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0077] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0078] In the embodiments of the above-described device, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0079] Through the above embodiments, this invention acquires target reference images and full monitoring videos of abnormal waste through a visual monitoring system, preprocesses the target reference images to generate target feature vectors, uses a target detection model to identify and extract waste disposal events from the full monitoring videos, arranges the waste disposal events in ascending order according to the waste disposal timestamps, and generates an original sequence of waste disposal events. Based on the original sequence of waste disposal events, it compares and analyzes steady-state image frames using an inter-frame difference algorithm, filters incremental object images and extracts incremental object feature vectors, and constructs a stacked temporal feature index library. It traverses the stacked temporal feature index library in reverse order and uses an early stopping mechanism to perform feature matching between incremental object feature vectors and target feature vectors to generate target disposal tracing results. This enables accurate indexing of abnormal waste disposal events and rapid tracing of disposal personnel, improving the efficiency and accuracy of abnormal waste disposal tracing.

[0080] This invention identifies and extracts garbage disposal events from a full suite of surveillance videos using a target detection model. It arranges these events in ascending order of timestamps, generating a raw sequence of events and precisely filtering out meaningless still images to reduce data redundancy. Furthermore, by utilizing a stacked temporal feature index library and leveraging the physical principle of garbage accumulation (last-in, first-out), it binds incremental object feature vectors to the disposal time dimension. This addresses the shortcomings of existing automated analysis techniques, such as neglecting scene physics and inefficient retrieval strategies, enabling efficient storage and rapid retrieval of feature data. Finally, this invention employs reverse depth... The priority detection strategy and early stop mechanism prioritize matching recent high-probability waste disposal events, avoiding the waste of computing resources caused by traversing the entire monitoring video frame by frame. This solves the problems of time-consuming, missed, and misjudged searches caused by the traditional needle-in-a-haystack approach to tracing abnormal waste disposal, significantly improving the response speed of abnormal waste disposal tracing. This invention automates the entire process from image acquisition, event extraction, stacked indexing to disposal tracing, without the need for manual screening, greatly reducing labor costs. At the same time, through precise comparison of incremental object feature vectors and target feature vectors, the reliability of tracing results is improved, which is conducive to the refined management of urban waste classification.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for tracing the source of abnormal garbage disposal based on computer vision stacked indexing, characterized in that, The method includes: The system acquires target reference images and full monitoring videos of abnormal waste through a visual monitoring system, and preprocesses the target reference images to generate target feature vectors. A target detection model is used to identify and extract garbage disposal events from the full amount of monitoring video. The garbage disposal events are arranged in ascending order according to the garbage disposal timestamp to generate the original sequence of garbage disposal events. Based on the original sequence of the garbage disposal event, the steady-state image frames are compared and analyzed by the inter-frame difference algorithm, incremental object images are filtered and incremental object feature vectors are extracted, and a stacked temporal feature index library is constructed. The stacked temporal feature index library is traversed in reverse order, and the incremental object feature vector is matched with the target feature vector through an early stopping mechanism to generate target delivery tracing results.

2. The abnormal garbage disposal tracing method based on computer vision stacked indexing according to claim 1, characterized in that, The process of acquiring a target reference image and full monitoring video of abnormal waste through a visual monitoring system, and preprocessing the target reference image to generate a target feature vector, specifically includes: The visual monitoring system acquires target reference images of abnormal waste from multiple angles in real time and uploads them to a smart terminal. The smart terminal automatically records the acquisition timestamp, abnormal waste collection point information, and acquisition angle information of the target reference image. Based on the collection timestamp, the smart terminal retrieves the full monitoring video through the SDK interface of the visual monitoring system according to the abnormal garbage collection point information; The target reference image is processed sequentially using the gray world algorithm for color balancing, bilinear interpolation for scaling, and Gaussian blur algorithm for noise suppression, resulting in a standard target reference image. The local texture features, global contour features, and color distribution features of the standard target reference image are extracted by the shallow, medium, and deep convolutional layers of the ResNet model, respectively. The local texture features, global contour features, and color distribution features are then subjected to numerical vector mapping and normalization using linear transformation and activation functions to generate the target feature vector.

3. The abnormal garbage disposal tracing method based on computer vision stack indexing according to claim 1, characterized in that, The process of identifying and extracting garbage disposal events from the full volume of surveillance video using a target detection model specifically includes: The full monitoring video is input into the target detection model. The video frame decoding module based on the target detection model parses the full monitoring video frame by frame into a continuous image frame sequence. The bounding box regression algorithm is used to identify the garbage containment image frame sequence in the continuous image frame sequence, and the garbage containment image frame sequence is extracted by pixel mask counting technology. A preset image pixel fluctuation threshold and fluctuation duration T are defined. Based on the start and end timestamps of the image pixel fluctuations in the waste containment image frame sequence, if the rate of change of the image pixel fluctuations in the waste containment image frame sequence is lower than the image pixel fluctuation threshold and the image pixel fluctuation duration is not lower than the fluctuation duration T, then the waste containment image frame sequence is determined to be in a stable state. If the rate of change of image pixels in the waste containment image frame sequence is higher than the image pixel fluctuation threshold, then the waste containment image frame sequence is determined to be in a changed state, and the waste containment image frame sequence is marked as the waste disposal event; After determining that the waste containment image frame sequence is in a changed state, if the rate of change of the image pixel fluctuation of the waste containment image frame sequence is lower than the image pixel fluctuation threshold, and the duration of the image pixel fluctuation is not lower than the duration of the fluctuation T, then it is determined that the waste containment image frame sequence has completed the change and returned to a stable state.

4. The abnormal garbage disposal tracing method based on computer vision stack indexing according to claim 3, characterized in that, The step of arranging the waste disposal events in ascending order according to the waste disposal timestamp to generate the original sequence of waste disposal events specifically includes: The garbage disposal event and the corresponding image frame number range are uploaded to the smart terminal for structured encapsulation to obtain a structured garbage disposal event; The smart terminal extracts the start timestamp of the waste disposal event and converts it into a standard start time value. Using the standard start time value as the sorting basis, the structured waste disposal events are sorted in ascending order using a timestamp sorting algorithm to generate the original sequence of the waste disposal events.

5. The abnormal garbage disposal tracing method based on computer vision stacked indexing according to claim 1, characterized in that, The process of analyzing steady-state image frames based on the original sequence of the garbage disposal event using an inter-frame difference algorithm, filtering incremental object images, and extracting incremental object feature vectors specifically includes: Based on the image frame number corresponding to the garbage disposal event and the image frame rate of the full monitoring video The range of old steady-state image frame numbers is obtained as follows: The range of new steady-state image frame numbers is ; Traverse the range of old steady-state image frame numbers in reverse order The last frame whose image pixel fluctuation rate is lower than the image pixel fluctuation threshold is selected as the old steady-state image frame, and the number range of the new steady-state image frames is traversed in ascending order. The first frame whose image pixel distribution volatility is lower than the image pixel volatility threshold is selected as the new steady-state image frame; A preset pixel difference threshold is set, and the old steady-state image frame and the new steady-state image frame are compared at the pixel level using the inter-frame difference algorithm to calculate the pixel difference. If the pixel difference is greater than or equal to the pixel difference threshold, the pixel corresponding to the pixel difference in the new steady-state image frame is retained to generate a difference image. The difference image is then subjected to binarization and morphological filtering in sequence, and the incremental object image is filtered using a connected component filtering algorithm. The incremental object image is input into a convolutional neural network model for feature encoding. A sliding filter is used to capture the local visual features of the incremental object image to generate a multi-channel feature map. A pooling layer is used to compress the size of the multi-channel feature map to generate a high-channel feature map. A fully connected layer is used to map the high-dimensional feature map into a fixed-dimensional feature vector of the incremental object.

6. The abnormal garbage disposal tracing method based on computer vision stacked indexing according to claim 1, characterized in that, The construction of the stacked time-series feature index library specifically includes: Each incremental object feature vector is data-bound using metadata tags to form an index data item. The index data item includes the incremental object feature vector, the start timestamp of the garbage disposal event corresponding to the incremental object feature vector, and the image frame index of the garbage disposal event corresponding to the incremental object feature vector in the full-volume monitoring video. Using a key-value pair storage mode, the index data items are pushed into the stacked time-series feature index library in ascending order according to a preset stacked storage rule.

7. The abnormal garbage disposal tracing method based on computer vision stack indexing according to claim 6, characterized in that, The reverse traversal of the stack-based temporal feature index library, and the feature matching of the incremental object feature vector with the target feature vector through an early stopping mechanism, specifically includes: Based on the reverse depth-first search strategy, the incremental object feature vector U is read one by one from the top of the stacked temporal feature index library, and the cosine similarity between the incremental object feature vector U and the target feature vector V is calculated. The corresponding calculation formula is as follows: In the formula, This represents the dimension of the incremental object feature vector, and the dimension of the target feature vector is equal to that of the incremental object feature vector; This represents the i-th element of the incremental object feature vector; This represents the i-th element of the target feature vector; A pre-set confidence threshold is used to adjust the cosine similarity through the early stopping mechanism. Compare with the confidence threshold, if the cosine similarity If the confidence level is greater than or equal to the confidence threshold, the feature matching is determined to be successful. The start timestamp of the garbage disposal event corresponding to the incremental object feature vector is locked as the target disposal timestamp, and the image frame index of the garbage disposal event corresponding to the incremental object feature vector in the full monitoring video is locked as the target monitoring video index. Feature matching is then stopped.

8. The abnormal garbage disposal tracing method based on computer vision stack indexing according to claim 7, characterized in that, Based on the target deployment timestamp and the target surveillance video index, the target deployment tracing result is generated, specifically including: Based on the target delivery timestamp and the target monitoring video index, locate the corresponding full monitoring video, and use the target delivery timestamp as the cutoff timestamp to backtrack the full monitoring video according to a preset behavior cycle to generate a backtracking monitoring video; Human body recognition algorithms are used to identify human objects in the retrospective surveillance video, and facial feature images and body posture feature images of the human objects are extracted. Based on the target deployment timestamp, the target baseline image, the retrospective monitoring video, and the facial and body feature images of the human object are aggregated and associated to generate the target deployment tracing result.

9. An abnormal garbage disposal tracing system based on computer vision stacked indexing, applied to the abnormal garbage disposal tracing method based on computer vision stacked indexing as described in any one of claims 1-8, characterized in that, The system includes: The image acquisition and preprocessing module is used to acquire target reference images and full monitoring videos of abnormal garbage through the visual monitoring system, and to preprocess the target reference images to generate target feature vectors. The waste disposal identification module is used to identify and extract waste disposal events from the full monitoring video using a target detection model, arrange the waste disposal events in ascending order according to the waste disposal timestamp, and generate the original sequence of waste disposal events. The feature index construction module is used to compare and analyze steady-state image frames based on the original sequence of the garbage disposal event, filter incremental object images and extract incremental object feature vectors, and construct a stacked time-series feature index library. The target delivery tracing module is used to traverse the stacked time-series feature index library in reverse order, and perform feature matching between the incremental object feature vector and the target feature vector through an early stopping mechanism to generate target delivery tracing results.

Images