System and method for monitoring loose waste collection enclosures

Abstract

The invention relates to a system for monitoring a loose waste collection enclosure (30), the system comprising: an image acquisition camera; an image processing unit (100); characterised in that the processing unit comprises at least one module for detecting each collection enclosure present within the acquired images on the basis of a first trained machine learning model; a module for determining the type of waste present in each portion of an image detected as being a collection enclosure on the basis of a second trained machine learning model; a module for calculating a fill level of each collection enclosure by analysing the enclosure images detected by the enclosure detection module.

Description

[0001] DESCRIPTION

[0002] TITLE OF THE INVENTION: SYSTEM AND METHOD FOR MONITORING WASTE COLLECTION ENCLOSURES

[0003] IN BULK

[0004] Technical field of the invention

[0005] The field of the invention is that of industrial and commercial waste management, in particular that of monitoring and tracking industrial and commercial waste collection enclosures. The invention relates in particular to a system and method for monitoring bulk waste collection enclosures, in particular industrial and commercial waste.

[0006] Technological background

[0007] Economic and industrial activities generate more than 59,400 tonnes of waste every minute worldwide, which represents more than 90% of the waste generated.

[0008] The problem of collecting and managing this industrial and commercial waste therefore presents a major challenge, both economic and environmental, and the operational collection challenges are numerous.

[0009] Field studies have shown that 80% of industrial and commercial waste collection bins are overloaded or underloaded, which in both cases presents specific problems, particularly related to the transport of these bins. Indeed, current regulations impose tonnage limits for waste transport, which are often exceeded in practice when the bins are overloaded. When bins are underloaded, their transport generates unnecessary travel, which contributes to increasing the costs and carbon footprint of their collections.

[0010] Regardless of the problems of transporting overloaded or underloaded skips, field studies show that 50% of the waste collected is poorly sorted or contaminated by other materials.

[0011] There is therefore a need for a solution to monitor industrial and commercial waste collection areas so that transport vehicles can be called upon only when the bins are filled to their nominal collection level and / or to detect any sorting defects in the collection area.

[0012] Throughout the following, the term "waste collection enclosure" refers to a space for storing bulk waste before its transport to a reprocessing, recycling or incineration center. These collection enclosures are generally arranged in the vicinity of the waste production or collection site (factory, warehouse, workshop, waste disposal center, etc.). By way of non-limiting example, a collection enclosure within the meaning of the invention is a waste container, a collection bin, a cell in a waste disposal center, a collection pit in a waste storage center, etc., the contents of which (and possibly the container) must be collected by a transport vehicle when this enclosure is full. A waste collection enclosure is therefore a delimited space intended to receive waste.

[0013] Throughout the text, waste collection zone means a geographical area comprising one or more waste collection areas within the meaning of the invention.

[0014] There are already solutions that allow you to monitor the fill level of a waste container.

[0015] One of these known solutions uses an ultrasonic sensor placed on the container and oriented towards the bottom of the container so as to be able to measure the distance between the sensor and the waste stored in the container. As soon as the measured distance is less than a predetermined distance, the container is considered full and its collection is organized. This solution is not adaptable to open collection areas, such as recycling center cells or open bins. In addition, this solution does not allow the type of waste collected to be detected and therefore any collection defects to be detected. This solution also proves to be ineffective if the container is not filled uniformly. In particular, if a large piece of waste is stored under the sensor, the system will consider the container to be full while the rest of the container may be free of waste.Finally, this solution requires equipping all containers, which can be costly and impractical, particularly when the waste producer rents collection bins from an external service provider, the latter not necessarily having planned to equip all these bins with such a solution.

[0016] Document US 10943356 describes another known solution implementing an image acquisition camera instead of the ultrasonic sensor. This camera is housed in a container and associated with image analysis and processing software. Like the previous solution, one of the disadvantages of this solution is that it is not adaptable to collection areas such as waste collection centers or open bins. In addition, this solution does not allow the type of waste collected to be detected and therefore any collection defects to be detected. Finally, this solution requires equipping all containers, which can be costly and impractical, particularly when the waste producer rents collection bins from an external service provider, the latter not necessarily having planned to equip all these bins with such a solution.

[0017] The inventors therefore sought to develop a new solution to overcome at least some of the drawbacks of known solutions.

[0018] Objectives of the invention

[0019] The invention aims to provide a system and a method for monitoring bulk waste collection enclosures, such as industrial or commercial waste.

[0020] The invention aims in particular to provide, in at least one embodiment, a system and a method which make it possible to automatically detect the filling level of the collection enclosures.

[0021] The invention aims in particular to provide a system and a method which make it possible to automatically detect the filling level of collection enclosures of all types, such as open waste containers, open collection bins, waste disposal cells, collection pits.

[0022] The invention also aims to provide, in at least one embodiment, a system and a method which make it possible to simultaneously monitor a collection area comprising several collection enclosures.

[0023] The invention aims in particular to provide, in at least one embodiment, a system and a method which make it possible to automatically detect the type of waste collected and their proportions in the collection enclosure.

[0024] The invention also aims to provide, in at least one embodiment, a system and a method which make it possible to alert operators in the field in the event of collection defects.

[0025] The invention also aims to provide, in at least one embodiment, a system and a method which make it possible to predict the times suitable for transporting waste from a collection enclosure.

[0026] The invention also aims to provide, in at least one embodiment, a system and a method which make it possible to improve the carbon footprint of an industrial or commercial activity generating waste.

[0027] The invention also aims to provide, in at least one embodiment, a system and a method which are easy to install and use.

[0028] Statement of the invention

[0029] To do this, the invention relates to a system for monitoring at least one bulk waste collection enclosure, such as a bin, a cell, a pit, comprising: an image acquisition camera configured to be able to acquire, at a predetermined time interval, an image of an area in which each collection enclosure to be monitored is arranged, a unit for processing the images acquired by said camera.

[0030] The system according to the invention is characterized in that said processing unit comprises at least the following modules: a module for detecting each collection enclosure present within said images acquired from a first machine learning model trained to be able to detect collection enclosures, this first learning module having been trained by means of a bank of training images, called enclosure banks, which included parts of images labeled as reflecting the presence of a collection enclosure and parts of images labeled as reflecting the absence of a collection enclosure, a module for determining the type of waste present in each part of the image detected by said enclosure detection module as being a collection enclosure, called enclosure images, from a second trained machine learning model,this second learning module having been trained using a training image bank, called a waste bank, which includes images of several types of waste likely to be collected in a collection enclosure, a module for calculating a filling rate of each collection enclosure by analyzing said enclosure images detected by said enclosure detection module.,

[0031] The system according to the invention therefore mainly comprises a camera and a unit for processing the acquired images.

[0032] The camera is configured (field of view and orientation) so that it can image an area that encompasses at least the monitored enclosure, or even several monitored enclosures.

[0033] The processing unit includes at least three separate modules for analyzing images acquired by the image acquisition camera.

[0034] All of the processing carried out by the three separate modules is based on the same images, i.e. on the images acquired by the camera of an area in which the enclosure(s) monitored by the system according to the invention are located, from the same point of view. No additional image provided by an additional sensor is necessary for the system according to the invention. Furthermore, no additional sensor, other than an image acquisition camera, is necessary for the system according to the invention.

[0035] The first module is a module for detecting the enclosures present in the images acquired by the camera. This module is therefore configured to extract from the images acquired by the camera sub-images each limited to a collection enclosure. These sub-images are designated throughout the text by the terminology of enclosure image. A collection area may comprise one or more enclosures and the image acquired by the camera may therefore comprise one or more enclosures. Thus, the first module can provide several images of enclosures from a single image acquired by the camera of the system according to the invention.

[0036] This first module implements a first machine learning model trained to detect collection enclosures within images. This learning model was trained from a training image bank (also referred to in the text as an enclosure bank). This enclosure bank consists of images of different types of enclosures (skips, cells, pits, etc.) and each image in this enclosure bank includes parts of images labeled (or tagged) as containing or not containing an enclosure. The parts of images labeled as reflecting the presence of a collection enclosure preferably include the boundaries of the collection enclosures. The machine learning model implemented by this first module can implement a supervised neural network.

[0037] The speaker images detected by the speaker detection module (also referred to as the first module in the text) are then processed by the second module of the processing unit.

[0038] This second module is a module for determining the type of waste present in each enclosure image. This module is therefore configured to determine the type of waste present in the enclosure imaged by the enclosure image processed by this module. This waste is, for example, made up of rubble, plastics, wood, cardboard, etc.

[0039] This second module implements a second machine learning model trained to detect the type of waste present in the image. This second learning module was trained from a training image bank (also referred to in the text as a waste bank) which includes a plurality of images of several types of waste likely to be collected in a collection enclosure. Each image in the waste bank is labeled with the waste visible in the image. The machine learning model implemented by this second module can implement a supervised neural network. This second module therefore provides an indication of the waste present in each enclosure image. The third module is a module for calculating the enclosure filling rate. This module is therefore configured to provide a filling level of the enclosure.Thus, as soon as the filling level reaches a predetermined value, a waste collection operation from the enclosure can be organized, for example by sending a notification to equipment or an operator.

[0040] Advantageously, the module for calculating the filling rate of each collection enclosure combines both the results of the analysis of the enclosure images detected by said enclosure detection module and the results of the type of waste identified by said waste type determination module.

[0041] Thus, the system allows, by knowing the type of waste present in the enclosure of the image provided by the second module, and by calculating the filling rate, to know the proportion of the different waste. Indeed, the system being configured to acquire images at predetermined time intervals, the invention makes it possible to follow the evolution of the filling of a container and the evolution of the filling by type of waste stored in the enclosure.

[0042] According to a variant of the invention, this third module implements a third machine learning model trained to detect the filling level.

[0043] This third learning model was trained using a training image bank, called a filling bank, which comprises a plurality of images of different enclosures containing different types of waste at different filling levels. Each image in the filling bank is labeled with a filling level between 0 and 100% filling. The machine learning model implemented by this third module, according to this variant, can implement a supervised neural network.

[0044] According to another variant of the invention, this third module calculates the filling rate by comparing at least one reference image corresponding to an empty enclosure with the image of said enclosure detected by said enclosure detection module. In other words and according to this variant, the third module does not implement a trained learning model, but a comparison of the enclosure image from the first module with one (or more) reference images corresponding to the empty enclosure.

[0045] The advantage of this variant is that it no longer depends on the training of a supervised neural network. Also, it does not depend on the labeling of training images, which is a human activity that can therefore vary depending on the person responsible for labeling the images in the training database. In addition, it allows us to bypass the training database (or padding bank) which must contain at least ten thousand images to be truly efficient.

[0046] Advantageously and according to this variant, said comparison of each reference image with said enclosure image comprises the detection and identification of similar key elements between said two images, called image descriptors.

[0047] According to this variant, the third module compares the images from image descriptors, which are the key points of the images. These descriptors are for example provided by a scale-invariant visual feature transformation algorithm, better known by the English acronym SIFT for "Scale-Invariant Feature Transform". The implementation of this algorithm makes it possible to identify the filling zones of the enclosure, independently of the differences in brightness, shading, and positioning of the enclosure from one image to another. The algorithm thus makes it possible to detect the key characteristics of the enclosure (for example the edges and edges of a container in the case of such an enclosure). The filling rate calculation module can then determine the height of the waste at a plurality of zones of the container by determining the variations detected at these characteristic points.

[0048] According to another variant, the filling rate calculation module combines at least in part the two previous variants, i.e. it implements a machine learning model trained to detect specific characteristics of the image (e.g. image descriptors), associated with a comparison of the image with a reference image.

[0049] Advantageously, the system according to the invention further comprises a solar panel connected to said image acquisition camera to be able to supply it with electrical energy.

[0050] The system in this variant is energy-independent and can be easily installed anywhere, regardless of the presence or absence of an electrical outlet.

[0051] Advantageously, a system according to the invention comprises a mast mounted in said area in which said collection enclosure is arranged, said mast carrying said image acquisition camera.

[0052] This variant allows the camera to be installed near the enclosure to be monitored. According to another variant, the camera can be mounted on a wall or structure present near the enclosure to be monitored.

[0053] The unit for processing the acquired images may be partially or totally embedded in the image acquisition camera or be formed partially or totally by a server remote from said camera. In the case where the processing unit is totally or partially formed by a remote server, the system advantageously comprises wireless communication means configured to transmit said images acquired by said camera to said processing unit.

[0054] Advantageously and according to this variant, said wireless communication means comprise at least 3G, 4G, 5G or WIFI connectivity.

[0055] This connectivity not only allows images to be transmitted from the camera to the processing unit, which is for example formed by a software application present on a remote server or a set of software applications and databases present on a set of remote servers (also referred to by the terminology of cloud), but also to transmit instructions to the camera.

[0056] Alternatively or in combination, it is also possible to implement wireless communication means that use low-speed connectivity such as a SigFox® or LoRa® network or any equivalent network, for the transmission of control data for example.

[0057] Advantageously and according to the invention, said predetermined time interval between two image acquisitions by said camera is a function of the change in the filling rate calculated by said filling rate calculation module and / or the type of waste determined by said waste type determination module.

[0058] The system according to this variant allows the image acquisition frequency to be adapted to the results provided by the processing unit. In particular, if the filling level increases rapidly, the acquisition frequency can be increased. Similarly, depending on the type of waste detected, it may be necessary to monitor the filling more regularly. Other rules for adapting the acquisition (and therefore processing) frequency can be provided depending on the monitoring objectives.

[0059] Advantageously, a system according to the invention further comprises a module for evaluating a mass and volume balance of the waste present in said enclosure.

[0060] According to this variant, the system can determine, from the results provided by the waste determination module and the filling rate calculation module, the mass and volume balance of the waste present in the enclosure. For this purpose, at each image analysis, the mass balance evaluation module retrieves information relating to the type of waste present and the filling level. The module can therefore determine the mass balance and the corresponding mass volume. The invention also relates to a method for monitoring at least one bulk waste collection enclosure, such as a skip, a cell, a pit, etc., said method comprising: an acquisition of images, at a predetermined time interval, of an area in which each monitored collection enclosure is arranged, a processing of the acquired images.

[0061] The method is characterized in that said image processing comprises: detecting each collection enclosure present within said acquired images by inputting the images into a first trained machine learning model, this first learning model having been trained using a bank of training images, called enclosure banks, which included parts of images labeled as reflecting the presence of a collection enclosure and parts of images labeled as reflecting the absence of a collection enclosure, the parts of images labeled as reflecting the presence of a collection enclosure comprising the boundaries of the collection enclosures, determining the type of waste present in each image part detected by said first learning model, by inputting each enclosure image into a second trained machine learning model,this second learning model having been trained using a training image bank, called a waste bank, which includes images of several types of waste likely to be collected in a collection enclosure, a calculation of a filling rate of each collection enclosure by analysis of said images of enclosures detected by said enclosure detection model.,

[0062] The technical advantages and effects of the system according to the invention apply mutatis mutandis to a method according to the invention.

[0063] The invention also relates to a system and a method, characterized in combination by all or part of the characteristics mentioned above or below.

[0064] List of figures

[0065] Other aims, characteristics and advantages of the invention will appear on reading the following description given solely for non-limiting purposes and which refers to the appended figures in which:

[0066] [Fig. 1] is a schematic view of a system according to one embodiment of the invention,

[0067] [Fig. 2] is a functional schematic view of a processing unit of a system according to an embodiment of the invention,

[0068] [Fig. 3] is a schematic view of an image acquired by a camera of a system according to an embodiment of the invention,

[0069] [Fig. 4] is a schematic view of the image extracts provided by a speaker detection module of a system according to an embodiment of the invention,

[0070] [Fig. 5] is a schematic view of an image of a bin intended to be processed by the modules for detecting the type of waste and calculating the filling rate of a system according to an embodiment of the invention,

[0071] [Fig. 6] is a schematic view of the images used by a waste type determination module of a system according to one embodiment of the invention,

[0072] [Fig. 7] is a schematic view of the images used by a module for calculating the filling rate of a system according to an embodiment of the invention.

[0073] Detailed description of an embodiment of the invention

[0074] In the figures, scales and proportions are not strictly adhered to for the purposes of illustration and clarity. In addition, identical, similar or analogous elements are designated by the same references in all figures.

[0075] Figure 1 illustrates a system according to an embodiment of the invention comprising a camera 10 for acquiring digital images of a waste storage bin 30 and a processing unit 100 for the digital images acquired by the camera 10. The images acquired by the camera 10 are transmitted to the processing unit 100 by a wireless communication network 50, which can be of any known type.

[0076] The embodiment described in connection with the figures comprises a processing unit formed by a server remote from the camera. That being said, nothing prevents, according to other embodiments, the processing unit from being partially or totally embedded on the image acquisition camera.

[0077] The camera 10 may be of any known type. It preferably has a wide angle or very wide angle (fish-eye camera) to be able to image a wide area including the storage bin 30 to be monitored. The camera is arranged at a distance from the bin, for example by being mounted on a structure, such as a post 12. This arrangement is made in such a way that the camera 10 overlooks the bin 30 and can take a picture of the bin 30 and the waste 32 stored in the bin. In other words, the camera 10 is oriented towards the bin 30 so as to be able to take pictures of the interior of the bin.

[0078] The system according to the invention can monitor several enclosures simultaneously provided that they can be accommodated in the area covered by the image camera.

[0079] Image acquisition is controlled by a control card housed in the camera 10. This control may, for example, consist of adjusting the magnification of the camera, the acquisition frequency, the aperture time and the ISO of the camera, etc.

[0080] The system further comprises a processing unit 100 for the images acquired by the camera 10. This processing unit 100 receives the images by wireless communication means 50 connecting the camera 10 and the processing unit 100.

[0081] In the embodiment of the figures, the processing unit is represented as a remote server connected to the image acquisition camera by wireless communication means. However, according to other embodiments not shown in the figures, the processing unit may be partially or totally embedded in the image acquisition camera.

[0082] This processing unit 100 comprises at least one processor, memories and software routines capable of implementing the different processing modules described below.

[0083] Throughout the following, a module is defined as a software element, a subset of a software program, which can be compiled separately, either for independent use or to be assembled with other modules of a program, or a hardware element, or a combination of a hardware element and a software subroutine. Such a hardware element may include an application-specific integrated circuit (better known by the acronym ASIC for the English term Application-Specific Integrated Circuit) or a programmable logic circuit (better known by the acronym FPGA for the English term Field-Programmable Gate Array) or a specialized microprocessor circuit (better known by the acronym DSP for the English term Digital Signal Processor) or any equivalent hardware or any combination of the aforementioned hardware. Generally speaking, a module is therefore an element (software and / or hardware) which makes it possible to perform a function.

[0084] The processing unit 100 comprises three main modules: a detection module 110 for the enclosures present in the acquired images, a determination module 120 for the type of waste present in each part of the image detected by said detection module 110, and a calculation module 130 for the filling rate of each enclosure.

[0085] Figure 2 is a functional diagram of the processing unit 100 implemented by a system according to the invention.

[0086] The processing unit 100 comprises, for example, a computing device 102 which must be understood in the broad sense (computer, plurality of computers, virtual server on the Internet, virtual server on the Cloud, virtual server on a platform, virtual server on a local infrastructure, server networks, etc.). This computing device typically comprises one or more processors 106, one or more memories 108 and a human-machine interface 104. The processing unit also comprises a database 116 for saving the results of the processing and for accessing information relating to previous processing. For example, the database 116 can store the data relating to the processing of the previously acquired images of the skip being processed, which makes it possible, for example, to determine the storage dynamics of the skip.

[0087] The processing unit 100 also comprises means for storing the trained learning models 118, 119 implemented by the invention. These means for storing the trained learning models may be servers separate from the equipment 102 or be saved within the equipment 102.

[0088] The computing device 102 further comprises the detection module 110, the determination module 120 and the calculation module 130. These modules use in particular the processors 106, the memories 108, the database 116 and the means for storing the trained learning models 118, 119, in order to be able to be executed.

[0089] The machine learning models implemented by the invention can be of different types. They can be a supervised learning neural network, a support vector machine (better known by the English acronym SVM for Support Vector Machine) or any other machine learning algorithm.

[0090] The module 110 for detecting the enclosures present in the images acquired by the camera 100 implements a trained machine learning model 118 to detect collection enclosures within the images. The learning base (or bank of enclosures) used to train the model consists of images of different types of enclosures (skips, cells, pits, etc.) following different orientations and different scales. Each image is labeled by a human operator by the type of enclosure it contains and by the delimitation of the enclosure within the image. The training of the model allows the latter to automatically determine the areas of presence of a collection enclosure within the image and the type of storage enclosures. The module 110 can then extract from the image of the skip 30 provided by the camera 100, the area of ​​the enclosure forming an image extract, called the enclosure image.This enclosure image can be used by the module 120 for detecting the type of waste present in the enclosure.

[0091] In the case where several speakers are monitored simultaneously by the same camera, the first module makes it possible to detect all the speakers monitored from a single image provided by the image acquisition camera.

[0092] The system according to the invention makes it possible to detect any type of enclosure arranged in the field of view of the camera. Also, if the enclosure is replaced by another or moved within the area imaged by the camera, the system continues to be functional, without needing to be informed of this change of enclosure or this movement of enclosure.

[0093] Note that according to one embodiment, this speaker detection module 110 can further be configured to determine the main dimensions of each detected speaker, in particular its width, its length and / or its height. To do this, the module can combine the knowledge of a known reference dimension within the image, obtained for example during a system calibration operation, and the knowledge of the field of view of the camera to deduce therefrom the dimensions of the speaker detected within the image. This reference dimension is for example the length of a wall adjacent to the collection area. This reference dimension makes it possible to deduce the area covered by a pixel of the camera. It is therefore possible to deduce the dimensions of the edges of the speaker from the number of pixels occupied by these edges within the image.Furthermore, the enclosures generally have standardized dimensions and the system has a database of the most common enclosures. Also, the module 110 can compare the measured dimensions with those present in the database to confirm the measurements and / or obtain the possible missing dimension (for example the height, from the measurement of the length and width).

[0094] The module 120 for determining the type of waste present in each image extract implements a trained machine learning model 119 to detect the waste present in the image extract detected by the module 110. The learning base (or waste bank) used to train the model consists of images of different types of waste (cardboard, wood, plastics, etc.) likely to be present in a storage enclosure.

[0095] The generation of the image bank was carried out by the applicant by placing image sensors on collection enclosures of different types (skip, cell, bulk storage on the ground) each containing one or more of the types of waste that one seeks to identify. A collection enclosure may comprise a single type of waste (for example, container dedicated to the collection of cardboard) or a plurality of waste (for example, general waste container from a waste disposal center which may contain plaster, glass wool, shredded or unshredded painted wood, etc.)

[0096] On each image of the image bank thus generated, the objects present are annotated separately using a segmentation approach. On each segmentation, a label is assigned. In other words, each object of each image is labeled by a human operator with the mention of the type of waste it represents. The training of the model allows the latter to automatically determine the contours of the objects present in the image and the types of waste present in the image.

[0097] The applicant has also enriched its learning database with the database known by the acronym TACO (available for example on the following website: httn: / / tacodataset.org). This image database contains a plurality of waste of all types in various environments. These images are manually labeled by operators. Other image databases can also be used to enrich the learning database, such as the databases known by the following English names: “Domestic Trash Dataset”; “WaDaBa”; “Open litter map”.

[0098] The applicant thus generated a learning base containing nearly 80,000 images. It should be noted, however, that it is possible to train learning models with a smaller number of images, depending on the quality of the images and the amount of waste present in each image.

[0099] According to an advantageous variant, the same image database is used for the enclosure detection module and for the waste type determination module. The generation of the image bank was carried out by the applicant by arranging image sensors of a plurality of collection enclosures of different types (skip, cell, bulk storage on the ground) each containing one or more of the types of waste that it is sought to identify.

[0100] On each image in the image bank thus generated, each speaker is annotated separately using a segmentation approach. On each segmentation, a label is assigned. In other words, each speaker in each image is labeled by a human operator with the indication of the speaker type.

[0101] Additionally, similar to the procedure described for the waste bank, the objects present in each image are annotated separately using a segmentation approach. A label is assigned to each segmentation. In other words, each object in each image is labeled by a human operator with the indication of the type of waste it represents.

[0102] As previously indicated, the machine learning models implemented by the invention may be supervised learning neural networks, support vector machines or any other machine learning algorithm. According to one embodiment, a convolutional neural network (known by the acronym CNN) is implemented for each learning model.

[0103] Learning is considered efficient if it allows the definition of a predictive model that adapts both to the training data and to new, unlabeled images. If the model does not adapt to the images in the training database, the model suffers from underfitting. If the model adapts too well to the images in the training database and is not able to correctly classify the new images, the model suffers from overfitting.

[0104] The learning phase aims to define the architecture of the neural network (number of layers, types of layers, learning steps, etc.) and the associated parameters (weightings of the layers and between the layers) which best model the different labels of the objects in the learning base and generate neither under-learning nor over-learning. This step of learning a neural network is known to those skilled in the art and the latter can refer to the existing literature to obtain more details on the steps to be implemented to do this. The module 130 for calculating the filling rate of the skip 30 can either implement another machine learning model trained from a learning base consisting of images of different enclosures having different fillings between 0 and 100%, or compare a reference image to the image of the skip being analyzed.

[0105] The image bank used for the detection module of each enclosure and for the waste type determination module can also be used for the filling rate calculation module, by labeling each enclosure with a filling rate.

[0106] Preferably, the system implements the comparison of the image extract with a reference image of the enclosure.

[0107] As described above, this comparison preferably implements a SIFT algorithm that identifies the fill areas of the enclosure, regardless of differences in brightness, shading, and positioning of the enclosure from one image to another. The algorithm thus makes it possible to detect key features of the enclosure (e.g., the edges and edges of a container in the case of such an enclosure). The fill rate calculation module can then determine the height of the waste at a plurality of areas of the container by determining the variations detected at these characteristic points.

[0108] According to another variant, the filling rate calculation module may combine at least in part the two previous embodiments. According to this variant, a machine learning model is trained to detect specific characteristics of the image, for example the edges of the speakers, and image comparison is used to compare the image with a reference image.

[0109] Figure 3 is a schematic representation of an image acquired by the camera 100 at a given time. This image comprises two skips 30a, 30b arranged in the vicinity of each other. The module 110 makes it possible to detect the two skips 30a and 30b and to extract the sub-images II, 12 which each comprise a skip monitored by the system according to the invention. Figure 4 schematically illustrates the two sub-images II and 12 respectively comprising the skips 30a and 30b extracted from the image acquired by the camera 100. The module 110 can also, according to one embodiment, be configured to anonymize the images II and 12 so as to guarantee compliance with the personal data protection regulation (better known by the acronym GDPR in France).

[0110] The steps illustrated in Figures 3 and 4 are implemented by the speaker detection module 110 in the image.

[0111] Each of these two sub-images II and 12 can then pass into the module 120 for determining the type of waste and into the module 130 for calculating a filling rate. These steps are illustrated schematically by figures 5, 6 and 7 in connection with the sub-image II of the bin 30a, it being understood that each sub-image of an enclosure of each image acquired by the camera 10 is processed in the same way by the different modules of the surveillance system according to the invention.

[0112] Figure 5 illustrates sub-image II of the bin 30a intended to be processed by the waste detection module 120 and by the module 130 for calculating a filling rate. The background of the image is deleted to keep only the bin 30a and the waste 32 stored in the bin.

[0113] It should be noted that the order of execution of modules 120 and 130 is not critical and that it is possible to first calculate the filling rate before determining the type of waste present in the bin 30a. However, in the following, it is expected that the determination of the type of waste is carried out before the calculation of the filling rate.

[0114] The module 120 is configured to detect the type of waste 32 present in the image. As previously indicated, the module 120 uses the trained machine learning model 119 which allows it to detect the types of waste present in the sub-image II of the bin 30a.

[0115] Figure 6 illustrates another aspect of the processing carried out by the module 120. The right-hand view is the state of the bin 30a being processed and the left-hand view is the state of the bin 30a during the previous acquisition of images of this same bin 30a. These previous results are for example stored in the database 116. Thus, the system can provide information on the variations in waste storage of the bin being analyzed. The system can therefore not only know the composition of the bin 30a at time / , but also know the filling dynamics and the filling times of this bin 30a. It is notably possible to know, in the event of detection of waste not admitted in the bin considered, the period during which this waste was added to the bin. The precision of the period will depend on the frequency of image acquisition by the camera.

[0116] To specifically assess the waste added between the two images, the second machine learning module can be implemented to determine whether the contents of the enclosure have changed and / or whether waste has been moved within the enclosure. It is then possible to use an algorithm, such as the SIFT algorithm described above, to more precisely detect the areas that have changed between two images, and cross-reference this information with the results of the analysis carried out by the machine learning module, to know precisely the waste that has been added between two images.

[0117] Figure 7 schematically illustrates the principle of calculating the filling rate of the bucket 30a according to one embodiment of the invention.

[0118] The module 130 retrieves from the database 116 a reference image Iref of the skip 30a being analyzed, this reference image Iref being an image of the same type of empty skip. This is the left image of Figure 7. The module 120 then calculates the filling level of the skip in the right image of Figure 7 by comparing this reference image Iref with the image of the skip 30a by implementing the SIFT algorithm previously described.

[0119] The invention is not limited to the embodiments described. In particular, a person skilled in the art will easily determine the possibilities offered by a system according to the invention from the various modules of the system. In particular, he will be able to evaluate the mass and volume balance of each skip, make predictions of skip filling periods from skip filling histories, etc. It is also possible to extend the waste detection and characterization routines to areas adjacent to the enclosures, such as waste stored at the feet of the enclosures. This can be achieved by widening the image analysis area to the surroundings of the enclosure. It is also possible to detect possible objects that block the removal of the enclosure (vehicle parked in front of a collection skip, trenches dug around the skip, etc.).To do this, it is necessary to enrich the waste database used by the waste detection module with images of objects likely to hinder access to the enclosure. It is also possible to create a specific learning database and implement an additional machine learning module specific to the analysis of the surroundings of storage enclosures.

[0120] The system has been described with an image acquisition camera in the visible range. However, according to a variant of the invention, a multispectral camera can be used to obtain images of the waste in frequency bands other than the visible range. In particular, an infrared camera can be used to obtain infrared images of the bulk waste to be detected.

Claims

CLAIMS System for monitoring at least one bulk waste collection enclosure (30, 30a, 30b), said system comprising: an image acquisition camera (10) configured to be able to acquire, at a predetermined time interval, an image of an area in which each collection enclosure (30, 30a, 30b) to be monitored is arranged, a processing unit (100) for the images acquired by said camera (10), characterized in that said processing unit (100) comprises at least the following modules: a detection module (110) for each collection enclosure (30a, 30b) present within said images acquired from a first automatic learning model (118) trained to be able to detect collection enclosures, this first learning module having been trained by means of a bank of training images, called banks of enclosures,which included parts of images labeled as reflecting the presence of a collection enclosure and parts of images labeled as reflecting the absence of a collection enclosure, a module (120) for determining the type of waste (32) present in each part of the image detected by said enclosure detection module as being a collection enclosure, called enclosure images, from a second trained machine learning model (119), this second learning module having been trained by means of a bank of training images, called a waste bank, which comprises images of several types of waste likely to be collected in a collection enclosure, a module (130) for calculating a filling rate of each collection enclosure by analyzing said detected enclosure images, by said enclosure detection module. System according to claim 1, characterized in that said filling rate calculation module (130) implements a third trained machine learning model, this third learning model having been trained by means of a training image bank, called a filling bank, which comprises images of different enclosures receiving different types of waste, each labeled with a filling level between 0 and 100% filling. System according to claim 1, characterized in that said filling rate calculation module (130) comprises the comparison of at least one reference image corresponding to an empty enclosure with said image of said enclosure detected by said enclosure detection module.System according to claim 3, characterized in that said comparison of said reference image with said enclosure image comprises the detection and identification of similar key elements between said two images, called image descriptors. System according to one of claims 1 to 4, characterized in that it further comprises a solar panel connected to said image acquisition camera (10) to be able to supply it with electrical energy. System according to one of claims 1 to 5, characterized in that it further comprises a mast (12) mounted in said area in which said collection enclosure is arranged, said mast carrying said image acquisition camera (10).System according to one of claims 1 to 6, characterized in that said predetermined time interval between two image acquisitions by said camera (10) is a function of the evolution of the filling rate calculated by said filling rate calculation module and / or the type of waste determined by said waste type determination module. System according to one of claims 1 to 7, characterized in that it further comprises a module for evaluating a mass and volume balance of the waste present in said enclosure. System according to one of claims 1 to 8, characterized in that said unit. processing unit is formed by a server remote from said camera and in that it further comprises wireless communication means (50) configured to transmit said images acquired by said camera to said processing unit. System according to claim 9, characterized in that said wireless communication means (50) comprise at least 3G, 4G, 5G or WIFI connectivity. Method for monitoring at least one bulk waste collection enclosure (30, 30a, 30b), said method comprising: an acquisition of images, at a predetermined time interval, of an area in which each monitored collection enclosure is arranged, a processing of the acquired images, characterized in that said image processing comprises: a detection of each collection enclosure present within said acquired images by inputting the images into a first trained machine learning model,this first learning model having been trained using a bank of training images, called enclosure banks, which included parts of images labeled as reflecting the presence of a collection enclosure and parts of images labeled as reflecting the absence of a collection enclosure, the parts of images labeled as reflecting the presence of a collection enclosure comprising the boundaries of the collection enclosures, a determination of the type of waste present in each part of the image detected by said first learning model, by inputting each enclosure image into a second trained machine learning model, this second learning model having been trained using a bank of training images, called a waste bank, which includes images of several types of waste likely to be collected in a, collection enclosure, a calculation of a filling rate of each collection enclosure by analyzing said enclosure images detected by said enclosure detection model.

Images