Household portrait construction and privacy protection optimization method and system based on garbage image recognition

Abstract

The application discloses a household portrait construction and privacy protection optimization method and system based on garbage image recognition, which comprises the following steps: collecting garbage images and corresponding positions in public areas, detecting and covering face areas and text areas locally on community service robots to obtain privacy processing images; identifying the category and volume interval of garbage targets, and quantifying the position and collection time into a position grid and a time slice to encapsulate an anonymous garbage event unit; uploading the anonymous garbage event unit to a property platform server, aggregating according to the position grid and the time slice, correcting the aggregation result according to the effective inspection times, forming garbage category distribution and unit inspection load; and combining the group service characteristics and the available capacity of the collection container to generate a service task. The application can reduce the risk of household privacy leakage and improve the accuracy of garbage load statistics and property service task generation.

Description

Technical Field

[0001] This application relates to the field of smart community management technology, specifically to a method and system for constructing resident profiles and optimizing privacy protection based on garbage image recognition. Background Technology

[0002] With the advancement of smart community construction, property management is gradually adopting community service robots, fixed camera equipment, and image recognition technology to inspect areas such as public roads, garbage disposal points, and public corridors to identify garbage categories, estimate garbage quantities, and arrange collection operations accordingly. Current technologies typically upload inspection images to a server for centralized identification and statistical analysis, or directly generate collection tasks based on the results of a single image recognition. Since images of public areas may contain faces, house numbers, resident identification, and other sensitive information, the transmission and long-term storage of raw images can easily increase the risk of privacy leaks. Furthermore, the precise location, specific collection time, and image content can also indirectly link to the activities of specific residents.

[0003] Furthermore, there are discrepancies in the number of inspections, inspection duration, and operational status of community robots. Existing statistical methods often directly reflect the regional waste load by counting the number of identified waste targets or estimating their volume, without fully considering the impact of actual inspection coverage in different areas and time periods. This can easily lead to misinterpreting differences in inspection frequency as differences in waste generation patterns. For the same waste target appearing repeatedly in continuous video feeds, the lack of an effective duplicate filtering mechanism can result in duplicate counting. Consequently, the resulting statistical results are difficult to accurately represent the waste category distribution and service needs in different public areas and time periods, and are also detrimental to the rational determination of collection timing, adjustment of inspection periods, and allocation of sorting and disposal facilities. Therefore, how to improve the statistical reliability of waste inspection data and the accuracy of property service task generation while reducing image privacy risks has become a technical problem that needs to be solved in this field. Summary of the Invention

[0004] This application provides a method and system for constructing resident profiles and optimizing privacy protection based on garbage image recognition, so as to at least solve some of the technical problems existing in the related technologies described above.

[0005] According to a first aspect of the embodiments of this application, a method for constructing resident profiles and optimizing privacy protection based on garbage image recognition is provided, including: Collect images of trash in public areas and obtain the location of the public area corresponding to the time of collection; The community service robot locally detects face and text regions in the public area garbage image, and performs overlay processing on the face and text regions to obtain a privacy-processed image; The privacy-processed image is subjected to garbage target identification to obtain the garbage category and volume range of each garbage target; the location of the public area is quantized into a public area location grid, the acquisition time is quantized into a time slice, and the garbage category, volume range, public area location grid and time slice of each garbage target are encapsulated into an anonymous garbage event unit; The anonymous garbage event units are uploaded to the property management platform server and aggregated according to the public area location grid and time slice. The effective number of inspections is determined based on the inspection task completion record, and the aggregation result is corrected based on the effective number of inspections to form community resident group service characteristics including garbage category distribution and unit inspection load. Service tasks are generated based on the service characteristics of the community residents and the available capacity of the corresponding collection containers.

[0006] As an optional solution, the step of collecting public area garbage images and obtaining the public area location corresponding to the collection time includes: a camera collecting red, green and blue image frames, a positioning unit outputting public area location samples with positioning time, and an edge processor pairing the red, green and blue image frames with the public area location sample with the smallest time difference; when the time difference exceeds a preset value, the red, green and blue image frames are discarded, and when the time difference does not exceed the preset value, the pairing result is written to the first partition of the local cache unit.

[0007] As an optional approach, the detection and overlay processing of face and text regions in the public area garbage image includes: scaling and normalizing the red-green-blue image frame and inputting it into a quantized YOLO-tiny dual-scale face detection network to obtain the face region; converting the red-green-blue image frame into a grayscale image and obtaining the text region based on the horizontal gradient, binarization, and closing operation; expanding the boundaries of the face region and the text region; performing Gaussian blur on the expanded face region; and filling the expanded text region with pixels using the mean of the red, green, and blue channels of its outer ring.

[0008] As an optional approach, the garbage target identification includes: converting the privacy-processed image into a channel-first format and inputting it into an SSD detection network with MobileNet as the backbone network; outputting garbage category confidence and bounding box offsets from feature maps at multiple scales via a classification head and a regression head; performing bounding box decoding, confidence filtering, and non-maximum suppression on the garbage category confidence and bounding box offsets to obtain the garbage category and bounding box of each garbage target; reading the depth value corresponding to the center of the bounding box, and determining the volume range only for garbage targets located within the volume calibration distance range.

[0009] As an optional approach, quantizing the location of the public area into a public area location grid and quantizing the collection time into a time slice includes: converting latitude and longitude coordinates into local planar coordinates based on the selected origin of the community map; rounding the local planar coordinates according to the preset grid side length to obtain the public area location grid; and truncating the collection time according to a preset fixed duration, using midnight of the day as the time reference, to obtain a time slice.

[0010] As an optional approach, before encapsulating the anonymous garbage event unit, duplicate targets are filtered out from garbage targets in consecutive processing frames; garbage targets with the same garbage category and a bounding box intersection-union ratio not lower than a preset value in adjacent processing frames are determined to be the same garbage target, and the anonymous garbage event unit is encapsulated only when the same garbage target appears for the first time, and the corresponding temporary state is cleared after the same garbage target has not appeared for a preset number of consecutive processing frames.

[0011] As an optional solution, the local cache unit includes a first partition and a second partition. The first partition stores the original image in a cyclical manner according to the first-in-first-out rule, and the second partition temporarily stores the anonymous garbage event units in the order of writing. When the property platform server connection is available, the anonymous garbage event units are LZ4 compressed and uploaded via MQTT messages. The message identifier and payload verification result returned by the property platform server are received. When the payload verification passes, the corresponding anonymous garbage event unit in the second partition is deleted.

[0012] As an optional solution, the property management platform server counts the number of anonymous garbage events within each volume interval according to the public area location grid, time slice, and garbage category. It determines the estimated volume of the aggregation unit based on the representative volume of each volume interval and determines the unit inspection load based on the effective inspection count. Using a continuous preset duration as the statistical period, it forms a garbage category distribution based on the unit inspection load of each garbage category and determines the high-load time slice based on the median, interquartile range, and 75th percentile of the unit inspection load of each time slice. The garbage category distribution, the unit inspection load, and the high-load time slice constitute the service characteristics of the community resident group.

[0013] As an optional approach, the service task generation includes: accumulating the estimated volume of each waste category within the current time slice, comparing the accumulated estimated volume with the available capacity of the corresponding collection container; when the accumulated estimated volume reaches the available capacity ratio threshold, generating a collection task including a public area location grid, a suggested arrival time slice, waste category, and estimated volume range; the available capacity ratio threshold is determined by statistically analyzing the historical overflow and invalid collection times of multiple candidate ratios, first selecting the candidate ratio with the fewest historical overflow times, and then selecting the candidate ratio with fewer invalid collection times from the tied candidate ratios.

[0014] According to a second aspect of the embodiments of this application, a resident profile construction and privacy protection optimization system based on garbage image recognition is also provided, comprising: This includes community service robots and property management platform servers; The community service robot includes: an image acquisition module, used to acquire images of garbage in public areas and obtain the location of the public area corresponding to the acquisition time; A privacy processing module is used to detect face regions and text regions in the public area garbage image locally by the community service robot, and to perform overlay processing on the face regions and the text regions to obtain a privacy-processed image; An anonymous event generation module is used to identify garbage targets in the privacy-processed image, obtain the garbage category and volume range of each garbage target, quantize the public area location into a public area location grid, quantize the acquisition time into a time slice, and encapsulate the garbage category, volume range, public area location grid and time slice of each garbage target into an anonymous garbage event unit. The data upload module is used to upload the anonymous spam event unit to the property management platform server; The property management platform server includes a feature generation module, which aggregates the anonymous garbage event units according to the public area location grid and time slice, determines the effective number of inspections based on the inspection task completion record, and corrects the aggregation result based on the effective number of inspections to form community resident group service features including garbage category distribution and unit inspection load. The service task generation module is used to generate service tasks based on the service characteristics of the community residents and the available capacity of the corresponding collection containers.

[0015] This application utilizes a community service robot to perform face and text region detection and coverage processing locally. It uses privacy-processed images as the sole input for waste target recognition, preventing the original images, facial features, text content, and detailed spatiotemporal information from entering the property management platform server, thus reducing the risk of resident privacy leaks and data misuse during public area inspections. Furthermore, waste categories, volume ranges, public area location grids, and time slices are encapsulated into anonymous waste event units, and duplicate targets in consecutive processing frames are filtered out, reducing statistical bias caused by repeated counting of the same waste target. The property management platform server corrects the aggregation results based on the effective number of inspections, mitigating the impact of inspection frequency, inspection duration, and equipment operating status differences on waste load statistics. This results in waste category distribution, unit inspection load, and high-load time slices that better reflect the actual service needs of the community's public areas. Based on the corrected group service characteristics and available collection container capacity, collection tasks are generated, helping to rationally determine collection timing, reduce waste overflow and ineffective collection, and provide a reliable basis for adjusting inspection periods and configuring classified disposal facilities.

[0016] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Furthermore, no embodiment in this disclosure is required to achieve all the effects described above. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0018] Figure 1 A flowchart illustrating a method for constructing resident profiles and optimizing privacy protection based on garbage image recognition, provided in this embodiment of the disclosure; Figure 2 A schematic block diagram of a resident profile construction and privacy protection optimization system based on garbage image recognition provided in this disclosure embodiment; Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0020] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, collected data, used data, generated data, processed data, etc.) and the data (including but not limited to data used for analysis, stored data, displayed data, collected information, used information, generated information, processed information, etc.) are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0021] This embodiment applies to scenarios where community service robots perform garbage inspections in residential roads, garbage disposal points, and public corridors. The implementing entities are the community service robot and the property management platform server. The community service robot is equipped with a camera, edge processor, positioning unit, communication module, and local caching unit. The property management platform server is equipped with a location grid aggregation module and a time-slice aggregation module. The edge processor locally performs face and text region processing, garbage target recognition, volume range determination, and anonymous garbage event unit encapsulation on the community service robot. The property management platform server receives anonymous garbage event units but does not receive raw images. Community resident group service characteristics refer to group service data statistically obtained by analyzing the location grid and time slices of public areas. This includes garbage category distribution, unit inspection load, and high-load time slices. It does not correspond to individual identities and does not include names, room numbers, door numbers, facial features, robot device serial numbers, or detailed locations that can be reconstructed from specific residents.

[0022] The community service robot can use a wheeled chassis with a cruising speed of 0.3 m / s to 0.8 m / s. The left eye of the binocular depth camera outputs red, green, and blue image frames, while the right eye outputs a depth map corresponding to the image at the time of acquisition and pixel coordinates. The edge processor uses an ARM Cortex-A72 quad-core processor and is connected to a privacy computing coprocessor running a quantized YOLO-tiny face detection network. The localization unit outputs latitude and longitude coordinates with the localization time. The communication module runs the Message Queuing Telemetry Transport (MQTT) protocol. The dual-partition local cache unit uses a first partition to cyclically store the original images, and a second partition to temporarily store anonymous garbage events.

[0023] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0024] Please see Figure 1, Figure 1 This is a flowchart of a method for constructing resident profiles and optimizing privacy protection based on garbage image recognition according to an embodiment of the present invention, as follows: Figure 1 As shown, the method includes the following steps: S1, collect images of garbage in public areas and obtain the location of the public area corresponding to the time of collection; S2, the community service robot locally detects the face region and text region in the public area garbage image, and performs overlay processing on the face region and the text region to obtain a privacy-processed image; S3, perform garbage target identification on the privacy-processed image to obtain the garbage category and volume range of each garbage target; quantize the public area location into a public area location grid, quantize the acquisition time into a time slice, and encapsulate the garbage category, volume range, public area location grid and time slice of each garbage target into an anonymous garbage event unit; S4, the anonymous garbage event units are uploaded to the property management platform server, aggregated according to the public area location grid and time slice, the effective number of inspections is determined according to the inspection task completion record, and the aggregation result is corrected according to the effective number of inspections to form community resident group service characteristics including garbage category distribution and unit inspection load; S5, generate service tasks based on the service characteristics of the community residents and the available capacity of the corresponding collection containers.

[0025] In some embodiments, for step S1, after the community service robot begins its patrol, the camera acquires red, green, and blue image frames at 15 to 30 frames per second, and the edge processor extracts the red, green, and blue image frames to be processed at a processing frequency of 5 frames per second. Each red, green, and blue image frame carries the acquisition time written by the edge processor, and the image resolution can be configured to 640 pixels by 480 pixels. Its data format is a three-dimensional array consisting of height, width, and three color channels. The positioning unit outputs public area location samples at 1 Hz to 5 Hz, and each public area location sample includes latitude and longitude coordinates and positioning time.

[0026] The edge processor calculates the time difference between the acquisition time of the red-green-blue image frame and the positioning time of the common area location sample, and pairs the common area location sample with the smallest time difference with the corresponding red-green-blue image frame. The calculation result of the time difference is taken as the absolute value, in milliseconds. When the time difference is less than 500 milliseconds, the edge processor establishes an association between the red-green-blue image frame, the acquisition time, and the common area location sample, and writes the association result to the first partition of the local cache unit. When the time difference exceeds 500 milliseconds, the corresponding red-green-blue image frame is discarded and not proceeded to subsequent processing. 500 milliseconds is determined by the cruise speed and the position grid size, corresponding to a displacement of 0.4 meters at a cruise speed of 0.8 meters per second. The image discard ratio and the number of cross-grid incorrect pairings are counted at 200 milliseconds, 500 milliseconds, and 1000 milliseconds, respectively. The 500-millisecond value with 0 cross-grid incorrect pairings and the lowest image discard ratio is selected.

[0027] As an example, the acquisition time of the red-green-blue image frame is 18:23:15.400. The positioning times of the two consecutive common area location samples are 18:23:15.000 and 18:23:16.000, respectively, with time differences of 400 and 600 milliseconds. The edge processor selects the previous common area location sample; when the closest time difference is 700 milliseconds, the corresponding red-green-blue image frame is discarded. The right eye depth map is associated with the retained red-green-blue image frame at the same acquisition time and temporarily stored in the edge processor's main memory for later reading of the depth value of the garbage target's center location.

[0028] In some embodiments, for step S2, the edge processor reads the red-green-blue image frames that have been paired with the public area location samples and passes them to the privacy computing coprocessor. The total input of the quantized YOLO-tiny face detection network is the red-green-blue image frames. Before inference, the 640-pixel by 480-pixel red-green-blue image frames are converted to 416-pixel by 416-pixel using proportional scaling and edge padding. Then, the pixel values ​​of each color channel are divided by 255 to form a four-dimensional input tensor with a batch size of 1, 3 channels, and a height and width of 416. During the scaling process, the horizontal scaling factor, the vertical scaling factor, and the edge padding width are recorded simultaneously. The face detection boxes output by the network are then reconstructed to the 640-pixel by 480-pixel red-green-blue image frame coordinate system.

[0029] The quantized YOLO-tiny face detection network consists of a backbone feature extraction section and a dual-scale detection section. The backbone feature extraction section sequentially uses 3x3 convolutional layers with 16, 32, 64, 128, 256, and 512 output channels, connected by 2x2 max-pooling layers. This is followed by a series of 3x3 convolutional layers with 1024 output channels, a 1x1 convolutional layer with 256 channels, and a 3x3 convolutional layer with 512 channels, forming a 13x13 scale detection feature map. The convolutional layers employ batch normalization and linear units with leakage correction.

[0030] The dual-scale detection part sets up a first prediction layer on a 13x13 scale detection feature map. Simultaneously, a branch is drawn from the aforementioned 1x1 convolutional layer with 256 output channels, passing through a 1x1 convolutional layer with 128 output channels and undergoing a 2x upsampling. This result is then concatenated with the 26x26 scale feature map in the backbone network channel-wise. The concatenated result is then passed through a 3x3 convolutional layer with 256 output channels to form a 26x26 scale detection feature map, which is then connected to the second prediction layer. Each grid position in each prediction layer is configured with 3 prior boxes. Each prior box outputs 4 bounding box parameters, 1 face presence probability, and 1 face class probability. The total network output is a face candidate box tensor at both scales. The privacy computation coprocessor performs coordinate restoration, confidence filtering, and non-maximum suppression on the candidate boxes to obtain the face regions, which are represented by rectangular boxes in the original red-green-blue image frame coordinate system.

[0031] For face confidence thresholds between 0.30 and 0.80, and non-maximum suppression intersection-union ratio (NUC) thresholds between 0.30 and 0.60, the validation set is iterated through. First, combinations with the fewest missed detections are selected, followed by combinations with fewer false positives and duplicate bounding boxes. For example, when both thresholds are set to 0.45, candidate boxes with a face presence probability of 0.82 and a face class probability of 0.91 have a confidence of 0.746 and participate in NUC; candidate boxes with a confidence of 0.31 are discarded.

[0032] The training data for the quantized YOLO-tiny face detection network consists of authorized community public area images and publicly available face images, with training samples labeled with rectangular bounding boxes. The input during training remains proportionally scaled, edge-padded, and channel-normalized red-green-blue image frames, and the output remains two-scale face candidate box tensors, consistent with the inference phase. The training loss is:

[0033] in, Indicates the total training loss; The mean square error representing the coordinates of the face center; The mean square error represents the square root of the width and height of a human face; The binary cross-entropy represents the probability of the presence of a face corresponding to a prior bounding box containing a face. The binary cross-entropy represents the probability of a face being present in a prior bounding box that does not contain a face. The binary cross-entropy represents the probability of face categories; each coordinate is divided by the image width or height before being used in the calculation. , , , and These represent the weights of the corresponding loss terms, and all values ​​are non-negative. Each weight is determined through a grid search on the validation set, prioritizing the number of missed face detections over the number of false positives; for example, the weights are 5, 5, 1, 0.5, and 1 respectively.

[0034] Training employs the Adam optimizer with an initial learning rate configurable to 0.001. The first five epochs are used for linear warm-up, followed by cosine annealing to reduce the learning rate to 0.00001. Training lasts for 80 to 150 epochs with batch sizes ranging from 16 to 32. Training stops when the validation loss fails to decrease for 15 consecutive epochs. After training, the weights and activation values ​​are quantized to 8-bit integers. If the face false negative rate of the quantized model increases by more than 1 percentage point compared to the unquantized model, calibration images are added and the model is requantized.

[0035] The text region includes rectangular areas in the red-green-blue image frames that display house numbers or resident identification characters. The edge processor converts the red-green-blue image frames into grayscale images, uses the horizontal Sobel operator to calculate the grayscale gradient of adjacent pixels in the horizontal direction, and then uses the inter-class variance maximization method to determine the binarization threshold, adjusting the gradient magnitude... Figure 2 The text is valued and a rectangular structuring element closing operation is performed to connect adjacent character strokes within the same text line into candidate connected regions. The edge processor calculates the aspect ratio, edge pixel ratio, and foreground pixel density of each candidate connected region, retaining the candidate connected regions that conform to the text region calibration range to obtain the text region.

[0036] The aspect ratio range and lower limit of edge pixel percentage for text regions are determined by traversing the labeled images, selecting the group with the fewest missed detections and the fewest false positives. For example, a candidate connected region with a width of 96 pixels, a height of 24 pixels, an aspect ratio of 4, an edge pixel percentage of 0.37, and falling within the calibration range is identified as a text region.

[0037] The edge processor expands the face and text regions outwards respectively. The horizontal expansion is 10% of the width of the corresponding rectangle, and the vertical expansion is 10% of the height of the corresponding rectangle. When the expanded coordinates exceed the boundaries of the red, green, and blue image frames, they are truncated to the image boundary. The 10% expansion ratio is calibrated using four candidate ratios: 5%, 10%, 15%, and 20%. The ratio with zero exposed pixels in the privacy region and minimal overlap with the garbage target annotation box is selected.

[0038] For the expanded face region, the privacy computing coprocessor uses a 15x15 pixel Gaussian kernel for blurring. For the expanded text region, the edge processor reads the 5-pixel ring around the text region, calculates the pixel mean values ​​for the red, green, and blue channels, and fills the text region with the corresponding channel mean values. When the face region overlaps with the text region, Gaussian blur is performed first, and then the text region filling result covers the overlapping area. The processed complete image is denoted as the privacy-processed image, and its resolution and number of channels remain 640 pixels by 480 pixels by 3 channels. The privacy-processed image is written to the edge processor's main memory via a direct memory access (DMA) channel, serving as the sole image input for garbage object recognition; face candidate boxes, text candidate boxes, face features, and text content are not written to the second partition.

[0039] For example, if the face region is defined as 120 to 180 pixels horizontally and 80 to 160 pixels vertically, after expansion, the horizontal region is increased by 6 pixels and the vertical region by 8 pixels. Then, a Gaussian blur is applied to the expanded region. When the mean values ​​of the red, green, and blue channels of the 5-pixel ring surrounding the text region are 118, 124, and 130 respectively, the edge processor uses the corresponding channel mean values ​​to cover all pixels within the text region. This coverage process preserves the garbage contours and scene background required for subsequent garbage object recognition in the privacy-processed image, while no longer transmitting face details and text character pixels to subsequent steps.

[0040] In some embodiments, for step S3, the edge processor converts the privacy-processed image into a channel-first format. The pixel values ​​of each color channel are divided by 255 and then subtracted from the mean value of the corresponding channel in the training set, forming a four-dimensional input tensor with a batch size of 1, 3 channels, a height of 480, and a width of 640. The SSD detection network, with MobileNet as its backbone, receives the four-dimensional input tensor corresponding to the privacy-processed image and outputs the garbage category confidence and bounding box offset at each scale. After bounding box decoding, confidence filtering, and non-maximum suppression, the garbage category and bounding box of each garbage target are obtained. The model input corresponds in content to the privacy-processed image output in S1. The bounding boxes in the model output define the positions of garbage targets in the privacy-processed image, and the garbage category indicates the category to which the garbage target within the bounding box belongs.

[0041] The backbone network first sets up a 3x3 convolutional layer with a stride of 2 and an output of 32 channels, then concatenates 13 groups of depthwise separable convolutions consisting of 3x3 channel-wise convolutions and 1x1 pointwise convolutions. The number of output channels in each group is 64, 128, 128, 256, 256, 512, 512, 512, 512, 512, 1024, and 1024, respectively. The stride of the 2nd, 4th, 6th, and 12th groups is 2, and the stride of the rest is 1. Each convolutional layer is connected to a batch normalization layer and a modified linear unit 6, outputting two levels of feature maps: 30x40 and 15x20.

[0042] The SSD additional feature layer uses four sets of dimensionality-reducing convolutions and stride convolutions to form 8x10, 4x5, 2x3, and 1x1 feature maps sequentially. These four levels of feature maps are connected in parallel with the 30x40 and 15x20 feature maps output by the backbone network to the classification head and regression head, respectively. Four prior boxes are set at each feature map position. The classification head outputs the background category and the confidence scores of five waste categories: kitchen waste, recyclables, hazardous waste, other waste, and bulky waste. The regression head outputs the bounding box offsets.

[0043] The confidence threshold for waste category is between 0.30 and 0.80, and the non-maximum suppression intersection-union (IUU) threshold is between 0.30 and 0.60. Parameters are selected by iterating through the validation set, based on average accuracy and the number of duplicate boxes, for example, 0.50 and 0.45 respectively. If the confidence score for a target is 0.78 and all other categories are below 0.20, it is identified as kitchen waste. If the IUU of two candidate boxes of the same category is 0.62, only the one with the higher confidence score is retained.

[0044] The model training data consists of privacy-processed images collected by a community service robot and subjected to the same coverage processing. Each training sample is labeled with a garbage category and bounding box. Both the training and inference phases take privacy-processed image tensors arranged according to batch size, number of channels, height, and width as input, and output the garbage category confidence and bounding box offset corresponding to the 6-level feature map.

[0045] Training loss is

[0046] in, This represents the total training loss for garbage target identification; The cross-entropy loss represents the background category and the five garbage categories; This represents the smoothed absolute value loss between the predicted bounding box offset and the labeled bounding box offset; Indicates the number of prior boxes that match the labeled garbage target; This represents the localization loss weights. The bounding box coordinates are divided by the image width and height before being used in the calculation. The values ​​are calibrated in increments of 0.25 between 0.5 and 2.0. The average accuracy and average intersection-union ratio of the bounding boxes are statistically verified for each candidate value. The value with both indicators in the top range is selected, for example, 1.

[0047] Training employs the Adam optimizer with an initial learning rate configurable to 0.0005. After the first five linear warm-up rounds, cosine annealing is used. Training lasts for 100 to 200 rounds with batch sizes ranging from 8 to 24, and incorporates random pruning, brightness adjustment, weight decay, and early stopping. The same consecutive inspection segment is not simultaneously included in the training and validation sets.

[0048] The edge processor reads the corresponding depth map pixel value temporarily stored in main memory based on the center coordinates of the bounding box of the garbage target. When the depth value is within the volume calibration distance range, the edge processor calculates the pixel area of ​​the bounding box of the garbage target and estimates the volume based on the volume conversion factor; when the depth value is not within the volume calibration distance range, no anonymous garbage event unit is generated for the garbage target, and it waits for subsequent red, green, and blue image frames for re-identification. The volume calibration distance range is taken as the closest to the farthest distance from which the relative error of the estimated volume of the standard object does not exceed 20%, and is fixed in the edge processor.

[0049] Let the coordinates of the top left corner and the bottom right corner of the garbage target bounding box be respectively... and The bounding box pixel area is

[0050] in, Indicates the first The bounding box pixel area of ​​each garbage target is expressed in pixels squared; the coordinates are expressed in pixels. The volume conversion factor is calibrated using known volume standard objects. The camera installation height, tilt angle, and image resolution are consistent with the actual inspection conditions. Multiple known volume standard objects are placed within the volume calibration distance range, and the bounding box pixel area and measured volume of each standard object are recorded. The volume conversion factor is then determined using the least squares method.

[0051] in, This indicates the volume conversion factor, expressed in liters per square pixel. Indicates the number of calibration samples; Indicates the first The bounding box pixel area of ​​a standard object, in pixels squared; Indicates the first The measured volume of each standard sample is given in liters. The estimated volume of the waste target is given in...

[0052] in, Indicates the first The estimated volume of each waste target is given, in liters. The edge processor divides the estimated volume into four ranges: less than 5 liters, 5 to 15 liters, 15 to 30 liters, and not less than 30 liters. The boundaries of the volume ranges are determined based on the collection container capacity and the volume distribution verified manually; no updates are performed when using the above four fixed volume ranges.

[0053] For example, if a garbage target has a bounding box width of 80 pixels and a height of 60 pixels, with a bounding box pixel area of ​​4800 pixels squared and a volume conversion factor of 0.0032 liters per pixel squared, then its estimated volume is 15.36 liters. The edge processor classifies it into the volume range of 15 to 30 liters. Another garbage target with an estimated volume of 4.6 liters is classified into the volume range of less than 5 liters. Subsequent anonymous garbage event units only write the volume range, not the bounding box coordinates, depth value, bounding box pixel area, volume conversion factor, or estimated volume.

[0054] The edge processor converts the latitude and longitude coordinates in the paired public area location samples into local planar coordinates based on the selected origin of the community map. Both the horizontal and vertical coordinates are in meters. The public area location grid refers to the regional unit determined by both the horizontal and vertical grid indices after the community's public area is divided according to fixed grid side lengths.

[0055] The grid index is calculated according to the following formula:

[0056] in, and These represent the horizontal grid index and the vertical grid index, respectively. and These represent the horizontal and vertical coordinates of the local plane, respectively, in meters; This indicates the grid side length, in meters. In this embodiment, the grid side length is 50 meters. The edge processor concatenates the horizontal and vertical grid indices according to a fixed byte order to obtain the common area location grid.

[0057] The 50-meter grid side length is determined from 25 meters, 50 meters, and 100 meters. The number of times the same deployment facility crosses a grid and the number of times adjacent deployment facilities are in the same grid are counted. The candidate value with the former being 0 and the latter being less frequent is selected. The grid side length is also not less than 5 times the radius of the circular probability of the positioning error. For example, when the local planar coordinates are 126 meters horizontally and 84 meters vertically, and the grid side length is 50 meters, the horizontal grid index is 2 and the vertical grid index is 1.

[0058] A time slice refers to a time interval obtained by dividing the same natural day into fixed durations. The edge processor uses the timestamp corresponding to 00:00:00 of the current day as the time base, with a time slice length of 3600 seconds and a time slice index of [index missing].

[0059] in, Indicates the time slice index; This represents the timestamp corresponding to the time of data collection, in seconds. This represents the timestamp corresponding to 00:00:00 on that day, in seconds; This indicates the time slice length in seconds. Time slice lengths are specified from 30 minutes, 60 minutes, and 120 minutes. A 60-minute slice is selected to differentiate waste changes across different time periods and ensures that each active public area location grid covers at least three inspection days within 28 days. When the data collection time is 18:23, the corresponding time slice index is 18.

[0060] Before encapsulating anonymous garbage event units, the edge processor performs duplicate target filtering on garbage targets in consecutive processing frames. For garbage targets of the same garbage category in adjacent processing frames, their bounding box intersection-union ratio is calculated:

[0061] in, Indicates the intersection-union ratio of the bounding boxes; This represents the pixel area of ​​the overlapping region of two bounding boxes, expressed in pixel squares. This represents the pixel area of ​​the merged region of two bounding boxes, expressed in pixel squares. When the bounding box intersection-union ratio (IU / U) is not less than 0.50, the two garbage targets are identified as the same garbage target, and an anonymous garbage event unit is encapsulated only upon their first appearance. When the IU / U is less than 0.50, they are treated as different garbage targets. After the same garbage target has not appeared for 5 consecutive processing frames, the edge processor clears its temporary state, and its reappearance is treated as a new garbage event.

[0062] The bounding box intersection-union ratio (IU / U) threshold and the number of frames to clear from the temporary state are determined by the labeled video. From candidate combinations of 0.30 to 0.70 and 3 to 10 frames, frames with the smaller sum of duplicate counts and missed counts (0.50 and 5 frames) are selected. For example, if the overlapping area of ​​the bounding boxes of the same type of garbage target in two adjacent frames is 3600 pixels squared, the merged area is 6000 pixels squared, and the IU / U is 0.60, the edge processor considers it the same garbage target. If the target is not identified again in consecutive frames 1 to 5, the temporary state is cleared after processing the 5th frame.

[0063] An anonymous waste event unit is a fixed-length structured record consisting of four business fields: waste category, volume range, public area location grid, and time slice. The waste category comes from the waste target identification result, the volume range comes from the range division result of the estimated volume, the public area location grid comes from the grid quantization result of the public area location sample, and the time slice comes from the time quantization result of the acquisition time. The edge processor uses fixed-length integer encoding according to a fixed field order, and the payload length of a single anonymous waste event unit does not exceed 256 bytes. The original image, bounding box coordinates, latitude and longitude coordinates, second-level timestamp, face detection result, text detection result, robot equipment serial number, and resident identifier are not written into the anonymous waste event unit.

[0064] For example, within the area corresponding to public area location grids 2 and 1, if food waste is identified in time slice 18 and its estimated volume falls within the range of 15 to 30 liters, then the anonymous waste event unit is sequentially written with the food waste category code, the 15 to 30 liter volume range code, the public area location grid code, and the 18-time-slice code. The edge processor writes the anonymous waste event unit to the second partition of the local cache unit. The red, green, and blue image frames and depth maps are still only stored in the first partition or main memory. When the first partition is full, the oldest original image is overwritten according to the first-in-first-out rule.

[0065] When the property management platform server connection is available, the edge processor reads the anonymous garbage event units in the order they were written, compresses them using LZ4 lossless compression, and forms MQTT messages. The message identifier, protocol version, payload length, and checksum fields are transmission control information and are not included in the anonymous garbage event units. After decompression and verification, the property management platform server returns the message identifier and payload verification results; if the verification passes, the corresponding anonymous garbage event unit in the second partition is deleted; if it fails or no acknowledgment is received, it is retained and retransmitted.

[0066] In some embodiments, for step S4, after the property management platform server receives the anonymous garbage event units, the location grid aggregation module merges and records them according to the public area location grid, and the time slice aggregation module continues to merge them according to time slice and garbage category within each public area location grid, and accumulates the number of anonymous garbage events in the four volume intervals respectively. The property management platform server sets a representative volume for each volume interval, which refers to the median measured volume of the manually verified sample in the corresponding volume interval, in liters; for open volume intervals of not less than 30 liters, the median of the manually verified volume of the interval in the most recent 90 days is read and updated once a month.

[0067] No. The grid of public area locations, the first The time slice and the first The estimated volume for each waste category is:

[0068] in, This indicates the estimated volume of the polymerization unit, in liters. Indicates the first The number of anonymous spam events within a given volume range; Indicates the first The representative volume for each volume range, in liters. For example, if the number of events in a certain polymer cell within the four volume ranges is 3, 2, 1, and 0 respectively, the representative volumes are 3 liters, 10 liters, 22 liters, and 40 liters respectively, resulting in an estimated volume of 51 liters.

[0069] The inspection task completion record is generated by the property management platform server after the robot completes the predetermined inspection path. It records the location grid of the public area, time slices, planned task duration, actual inspection duration, and the operating status of the camera, positioning unit, and edge processor. A valid inspection is counted when the actual inspection duration reaches 80% of the planned task duration, and the camera, positioning unit, and edge processor are in normal working order within the actual inspection duration; otherwise, it is not counted as a valid inspection. The 80% percentage is determined from candidate values ​​ranging from 70% to 95%, and the lowest percentage where increasing the candidate percentage by 5 percentage points results in a decrease of less than 5% in the number of missed inspections during manual review.

[0070] The property management platform server corrects the aggregated results based on the number of valid inspections. The unit inspection load is:

[0071] in, This indicates the unit of inspection load, expressed in liters per inspection. This indicates the estimated volume of the polymerization unit, in liters. This represents the number of effective inspections within the corresponding public area location grid and time slice, expressed in times. When the number of effective inspections is 0, the unit inspection load for the corresponding aggregated cell is not calculated, nor is it replaced by 0. Continuing with the previous example, if the estimated volume of the aggregated cell is 51 liters and the number of effective inspections is 3, then the unit inspection load is 17 liters per inspection.

[0072] The property management platform server uses a continuous 28-day statistical period to form service characteristics for community resident groups according to public area location grids. Waste category distribution is determined based on the cumulative unit inspection load of each waste category within the same public area location grid; the proportion of a particular waste category is the ratio of its cumulative unit inspection load to the cumulative unit inspection load of all waste categories. The property management platform server also calculates the median and interquartile range of unit inspection load by time slice, and defines time slices where the median unit inspection load is at or above the 75th percentile of the median of all time slices in that public area location grid as high-load time slices. The 75th percentile is defined from the 60th to the 90th percentile, first comparing historical overflow counts, then comparing the number of invalid tasks.

[0073] For example, if the median unit inspection load of a certain public area grid in time slots 18 to 21 is 12, 18, 25 and 15 liters per time respectively, and the 75th percentile is 20 liters per time, then time slot 20 is a high-load time slot; when the cumulative unit inspection load of each type of waste is 280, 120, 20, 70 and 10 liters per time respectively, the proportion of kitchen waste is 56% and the proportion of recyclables is 24%.

[0074] For short-term waste collection tasks, the property management platform server accumulates the estimated volume of each waste category within the current time slice and reads the available capacity of the collection containers corresponding to the public area location grid. Available capacity is the remaining capacity after subtracting the volume already recorded before the task started from the rated capacity of the collection container, in liters. When the accumulated estimated volume reaches the available capacity ratio threshold, the property management platform server generates a waste collection task. The task includes the public area location grid, suggested arrival time slice, waste category, and estimated volume range. If the accumulated estimated volume does not reach the available capacity ratio threshold, no waste collection task is generated, and the server continues to accumulate subsequently received anonymous waste event units within the same time slice.

[0075] The available capacity ratio threshold is set between 0.60 and 0.90, with intervals of 0.05. Historical collection records are read, and for each candidate ratio, the number of times the collection container overflows and the number of invalid collections where the actual volume collected upon arrival is less than 30% of the available capacity are counted. The candidate ratio with the fewest overflows is selected first, and then the candidate ratio with the fewest invalid collections is selected from the tied candidate ratios; for example, the available capacity ratio threshold is set to 0.75. When the available capacity of a collection container is 200 liters, the corresponding trigger volume is 150 liters. After the cumulative estimated volume in the current time slice reaches 152 liters, the property management platform server generates a collection task; when the cumulative estimated volume reaches 148 liters, it continues to wait for subsequent anonymous garbage event units.

[0076] For periodic service adjustments, the property management platform server compares high-load time slots and waste category distribution across two consecutive 28-day statistical periods. If the same high-load time slot occurs in the same public area location grid for two consecutive statistical periods, and the median unit inspection load for each slot is higher than the median daily average unit inspection load for that public area location grid, an inspection time slot adjustment suggestion is generated. This branch is executed only if both conditions are met; otherwise, no inspection time slot adjustment suggestion is generated. If the proportion of a certain waste category is at or above the 75th percentile of all waste categories in that public area location grid for two consecutive statistical periods, a corresponding waste disposal facility configuration suggestion is generated. This judgment and the inspection time slot adjustment suggestion are executed separately and do not overlap.

[0077] For example, if time slot 20 is a high-load time slot for two consecutive statistical periods, with median unit inspection loads of 24 liters and 27 liters per inspection respectively, exceeding the daily averages of 16 and 18 liters per inspection respectively, the property management platform server generates inspection time adjustment suggestions. If the proportion of food waste is 58% and 61% respectively, both at or above the 75th percentile, suggestions for configuring food waste disposal facilities are generated. After completing the collection task, service personnel enter the actual collection volume, whether overflow occurred, and the task completion time. The property management platform server updates the representative volume of the volume range monthly using the actual collection volume and re-calibrates the available capacity ratio threshold quarterly. Service feedback does not modify already formed anonymous waste event units.

[0078] In the above processing, the original images are only processed locally on the community service robot for face regions, text regions, and waste targets. The uploaded data is limited to waste category, volume range, public area location grid, and time slice. The property management platform server generates waste category distribution, unit inspection load, and high-load time slices based on the number of valid inspections, and generates collection tasks, inspection period adjustment suggestions, and disposal facility configuration suggestions accordingly. The above results are maintained at the granularity of public area location grid and time slice; face, house number, resident identification, and fine spatiotemporal information are not entered into the business fields of the property management platform server.

[0079] Please see Figure 2 , Figure 2 This is a structural block diagram of a resident profile construction and privacy protection optimization system based on garbage image recognition, provided in an embodiment of this application. Figure 2 As shown, the system includes a community service robot 20 and a property management platform server 30; The community service robot 20 includes: an image acquisition module 201, used to acquire images of garbage in public areas and obtain the location of the public area corresponding to the acquisition time; Privacy processing module 202 is used to detect face regions and text regions in the public area garbage image locally by the community service robot, and to perform overlay processing on the face regions and the text regions to obtain a privacy-processed image; Anonymous event generation module 203 is used to perform garbage target identification on the privacy-processed image, obtain the garbage category and volume range of each garbage target, quantize the public area location into a public area location grid, quantize the acquisition time into a time slice, and encapsulate the garbage category, volume range, public area location grid and time slice of each garbage target into an anonymous garbage event unit; Data upload module 204 is used to upload the anonymous spam event unit to the property management platform server; The property platform server 30 includes: a feature generation module 301, which is used to aggregate the anonymous garbage event units according to the public area location grid and time slice, determine the effective number of inspections based on the inspection task completion record, and correct the aggregation result based on the effective number of inspections to form community resident group service features including garbage category distribution and unit inspection load; The service task generation module 302 is used to generate service tasks based on the service characteristics of the community residents and the available capacity of the corresponding collection container.

[0080] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0081] Based on the same inventive concept, this application also provides an electronic device, the method corresponding to which can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. For example... Figure 3 As shown, Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods and / or technical solutions of the foregoing embodiments of the present application.

[0082] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processor, it performs the functions defined in the methods of this application.

[0083] Another embodiment of this application provides a storage medium storing computer program instructions thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application.

[0084] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A method for constructing resident profiles and optimizing privacy protection based on garbage image recognition, characterized in that, include: Collect images of trash in public areas and obtain the location of the public area corresponding to the time of collection; The community service robot locally detects face and text regions in the public area garbage image, and performs overlay processing on the face and text regions to obtain a privacy-processed image; The privacy-processed image is subjected to garbage target identification to obtain the garbage category and volume range of each garbage target; The location of the public area is quantized into a public area location grid, the collection time is quantized into a time slice, and the waste category, volume range, public area location grid and time slice of each waste target are encapsulated into an anonymous waste event unit; The anonymous garbage event units are uploaded to the property management platform server and aggregated according to the public area location grid and time slice. The effective number of inspections is determined based on the inspection task completion record, and the aggregation result is corrected based on the effective number of inspections to form community resident group service characteristics including garbage category distribution and unit inspection load. Service tasks are generated based on the service characteristics of the community residents and the available capacity of the corresponding collection containers.

2. The method according to claim 1, characterized in that, The process of collecting images of garbage in public areas and obtaining the location of public areas corresponding to the collection time includes: a camera collecting red, green, and blue image frames; a positioning unit outputting public area location samples with positioning time; and an edge processor pairing the red, green, and blue image frames with the public area location sample with the smallest time difference. If the time difference exceeds a preset value, the red, green, and blue image frames are discarded; if the time difference does not exceed the preset value, the pairing result is written to the first partition of the local cache unit.

3. The method according to claim 2, characterized in that, The process of detecting and covering face and text regions in the public area garbage image includes: scaling and normalizing the red-green-blue image frame and inputting it into a quantized YOLO-tiny dual-scale face detection network to obtain face regions; converting the red-green-blue image frame into a grayscale image and obtaining text regions based on horizontal gradient, binarization, and closing operations; expanding the boundaries of the face and text regions; performing Gaussian blur on the expanded face region; and filling the expanded text region with pixels using the mean of the red, green, and blue channels of its outer ring.

4. The method according to claim 3, characterized in that, The garbage target identification includes: converting the privacy-processed image into a channel-first format and inputting it into an SSD detection network with MobileNet as the backbone network; outputting garbage category confidence and bounding box offsets from feature maps of multiple scales through a classification head and a regression head; performing bounding box decoding, confidence filtering, and non-maximum suppression on the garbage category confidence and bounding box offsets to obtain the garbage category and bounding box of each garbage target; reading the depth value corresponding to the center of the bounding box, and determining the volume range only for garbage targets located within the volume calibration distance range.

5. The method according to claim 4, characterized in that, The step of quantizing the location of the public area into a public area location grid and quantizing the collection time into a time slice includes: converting latitude and longitude coordinates into local planar coordinates based on the selected origin point of the community plan map; rounding the local planar coordinates according to the preset grid side length to obtain the public area location grid; and truncating the collection time according to the preset fixed duration using midnight of the day as the time base to obtain a time slice.

6. The method according to claim 5, characterized in that, Before encapsulating the anonymous garbage event unit, duplicate targets are filtered out in consecutive processing frames; garbage targets with the same garbage category and a bounding box intersection-union ratio not lower than a preset value in adjacent processing frames are determined to be the same garbage target. The anonymous garbage event unit is encapsulated only when the same garbage target appears for the first time, and the corresponding temporary state is cleared after the same garbage target has not appeared for a preset number of consecutive processing frames.

7. The method according to claim 6, characterized in that, The local cache unit includes a first partition and a second partition. The first partition stores the original image in a cyclical manner according to the first-in-first-out rule, and the second partition temporarily stores the anonymous garbage event units in the order of writing. When the property platform server connection is available, the anonymous garbage event units are compressed using LZ4 and uploaded via MQTT messages. The message identifier and payload verification result returned by the property platform server are received. When the payload verification passes, the corresponding anonymous garbage event unit in the second partition is deleted.

8. The method according to claim 7, characterized in that, The property platform server counts the number of anonymous garbage events in each volume interval according to the public area location grid, time slice and garbage category, determines the estimated volume of the aggregation unit based on the representative volume of each volume interval, and determines the unit inspection load based on the effective inspection number. Using a continuously preset duration as the statistical period, a waste category distribution is formed based on the unit inspection load of each waste category. High-load time slots are determined based on the median, interquartile range, and 75th percentile of the unit inspection load of each time slot. The waste category distribution, the unit inspection load, and the high-load time slots constitute the service characteristics of the community resident group.

9. The method according to claim 8, characterized in that, The generated service task includes: accumulating the estimated volume of each type of waste within the current time slice, comparing the accumulated estimated volume with the available capacity of the corresponding collection container; when the accumulated estimated volume reaches the available capacity ratio threshold, generating a collection task including the public area location grid, suggested arrival time slice, waste type, and estimated volume range; the available capacity ratio threshold is determined by statistically analyzing the historical overflow and invalid collection times of multiple candidate ratios, first selecting the candidate ratio with the fewest historical overflow times, and then selecting the candidate ratio with fewer invalid collection times from the tied candidate ratios.

10. A resident profile construction and privacy protection optimization system based on garbage image recognition, characterized in that, This includes community service robots and property management platform servers; The community service robot includes: an image acquisition module, used to acquire images of garbage in public areas and obtain the location of the public area corresponding to the acquisition time; A privacy processing module is used to detect face regions and text regions in the public area garbage image locally by the community service robot, and to perform overlay processing on the face regions and the text regions to obtain a privacy-processed image; An anonymous event generation module is used to identify garbage targets in the privacy-processed image, obtain the garbage category and volume range of each garbage target, quantize the public area location into a public area location grid, quantize the acquisition time into a time slice, and encapsulate the garbage category, volume range, public area location grid and time slice of each garbage target into an anonymous garbage event unit. The data upload module is used to upload the anonymous spam event unit to the property management platform server; The property management platform server includes a feature generation module, which aggregates the anonymous garbage event units according to the public area location grid and time slice, determines the effective number of inspections based on the inspection task completion record, and corrects the aggregation result based on the effective number of inspections to form community resident group service features including garbage category distribution and unit inspection load. The service task generation module is used to generate service tasks based on the service characteristics of the community residents and the available capacity of the corresponding collection containers.

Images