Multi-dimensional seeding alarm method and system for seeding machine

Abstract

The invention relates to the technical field of seeding monitoring, and particularly discloses a multi-dimensional seeding alarm method and system for a seeder, and the method comprises the steps: collecting images of each seeding hole in the rotation process of a seed disc based on industrial camera equipment built in a seeding apparatus of the seeder; performing target identification on the image of each seed-metering hole, positioning a seed target and a contour feature thereof, and judging a seeding state; counting sowing states at different moments to obtain a state sequence, performing time domain analysis and feature analysis on the state sequence, and correcting the sowing states; counting the corrected seeding state of each seeding hole, constructing visual information in combination with an image, and feeding back the visual information to a main end of the seeding machine; according to the invention, a camera monitoring and alarm monitor combined device is adopted, so that a driver can see the seed suction condition on the seed plate of the seed-metering device through the camera only by observing the display screen, and the intelligent degree is extremely high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seeding monitoring, in particular to a multi-dimensional seeding alarm method and system for a seeding machine. BACKGROUND

[0002] The seeding machine is important operation equipment for realizing precise seeding of crops in agricultural production, and the seed discharge quality of the seeding machine directly affects the seedling emergence rate and field uniformity. In actual operation process, the seeding machine often uses photoelectric, mechanical or capacitive sensors to monitor the seed passing condition, and prompts the abnormal state such as missing seeding and double seeding according to the collected signals.

[0003] However, the existing seeding monitoring method mainly depends on a single sensing signal for judgment, which is easily disturbed by factors such as dust, soil particles, seed coating debris and vibration in complex operation environment, resulting in miscounting or false alarm, and affecting the reliability of seeding quality judgment. At the same time, it is difficult to distinguish the specific abnormal types such as no seed suction, double seed suction and foreign matter blockage by relying on the seed passing signal alone, which is not conducive to the timely adoption of targeted adjustment measures by the driver.

[0004] With the development of intelligent agricultural equipment, some seeding machines begin to introduce video monitoring means to improve the accuracy of seeding state recognition, but the video acquisition and recognition process consumes a lot of computing resources. If the video is analyzed continuously and frequently during the seeding process, the system cost and energy consumption will be significantly increased, which limits its practicability in agricultural field.

[0005] Therefore, a low-power comprehensive seeding state recognition scheme is needed to simply and efficiently recognize the seeding state in multiple dimensions, thereby improving the reliability and intelligent level of the seeding operation. SUMMARY

[0006] The purpose of the present application is to provide a multi-dimensional seeding alarm method and system for a seeding machine to solve the problems raised in the background.

[0007] To achieve the above purpose, the present application provides the following technical scheme: A multi-dimensional seeding alarm method for a seeding machine, the method comprising: Collecting images of each seed discharge hole in the seed disc rotation process based on an industrial camera device built into the seed discharge device of the seeding machine; wherein the image collection frequency of each seed discharge hole is different; Performing target recognition on the images of each seed discharge hole, locating the seed target and its contour features, and determining the seeding state; The sowing states at different time points are counted to obtain a state sequence, time domain analysis and feature analysis are performed on the state sequence, and the sowing state is corrected; the feature analysis includes space-time analysis to obtain instantaneous features and persistent features, the instantaneous features are used to represent the seed state in the image, and the persistent features represent the change of the seed state in a period of time, in the extraction of the persistent features, a current image is taken as a tail image to obtain an image sequence in a period of time, and a persistent feature extraction process is performed; The corrected sowing states of the seed holes are counted, visual information is constructed in combination with the images, and the visual information is fed back to the seeding machine terminal; wherein the visual information at least includes images in different sowing states and state statistical information of the seed holes.

[0008] As a further scheme of the application, the step of collecting the images of the seed holes in the seed disc rotation process by the industrial camera device built in the seed dispenser of the seeding machine comprises: The seed passing signal output by the seeding machine alarm monitor is collected; the seed passing signal is used to represent whether there is an object passing through the seed hole; The shooting instruction is generated based on the seed passing signal, and the shooting instruction is sent to the industrial camera device built in the seed dispenser of the seeding machine; The images in the seed holes in the seed disc rotation process are collected based on the industrial camera device, and the collected images are time-stamped; The time points of the seed passing signal are counted in real time, and the self-frequency of the industrial camera device is determined; the self-frequency is used to collect images at regular intervals when there is no shooting instruction.

[0009] As a further scheme of the application, the step of identifying the images of the seed holes, locating the seed target and its contour features, and determining the sowing state comprises: For any seed hole, the image at the latest time point is read in real time, and the image is preprocessed; the preprocessing includes image cropping, size normalization and noise suppression; The seed area is located in the preprocessed image, the seed area is segmented based on an image segmentation algorithm, and the seed target and its contour features in the seed hole are extracted; The sowing state is determined according to the extracted seed target and its contour features.

[0010] As a further scheme of the application, the step of counting the sowing states at different time points to obtain a state sequence, performing time domain analysis and feature analysis on the state sequence, and correcting the sowing state comprises: The sowing states at different time points are counted to obtain a state sequence; The state sequence is converted into a numerical sequence; A difference sequence of the numerical sequence is calculated, and a marked abnormal moment is determined according to the difference sequence; the abnormal moment refers to a moment point corresponding to an element sequence number when a difference absolute value reaches a preset threshold value; A seed target and a contour feature of the image at the abnormal moment are read, and a geometric feature is determined according to the seed target and the contour feature as a transient feature; wherein the transient feature includes a seed number, a hole proportion and a seed spacing; An image sequence with the image at the latest moment as a tail image is determined based on a preset time span, and a seed change vector is determined by performing difference identification on the image sequence as a continuous feature; Based on the extracted transient feature and the continuous feature, the current seeding state of the seed hole is determined as a corrected seeding state; the seeding state at least includes a normal seed suction state, a double seed suction state, a non-seed suction state and a foreign matter blockage state.

[0011] As a further scheme of the present application: the step of statistically correcting the seeding state of each seed hole and combining the image to construct visual information and feeding back to the total end of the seeding machine includes: Reading the state sequence and the difference sequence thereof; Segmenting the state sequence according to the difference sequence and synchronously determining the image display proportion in each time period; Randomly extracting images in different time periods based on the image display proportion and synchronously extracting images at abnormal moments; Taking the state sequence and the extracted images as visual information and feeding back to the total end of the seeding machine.

[0012] As a further scheme of the present application: the method further includes: Establishing a connection channel with a sensor built in the seeding machine, obtaining data of each sensor based on a preset data template, and obtaining a real-time matrix; Reading the seeding state at each moment and constructing a sample set of the real-time matrix to the seeding state; Training a pre-detection model based on the sample set, and recursively adjusting the self-frequency of the industrial camera device according to the error rate of the pre-detection model.

[0013] The technical scheme of the present application further provides a multi-dimensional seeding alarm system for a seeding machine, the system including: An image acquisition module for acquiring images of each seed hole in the seed disc rotation process based on an industrial camera device built in the seed dispenser of the seeding machine; wherein the image acquisition frequencies of the seed holes are different; A seeding state determination module for performing target identification on the images of each seed hole, locating a seed target and a contour feature, and determining a seeding state; The sowing state correction module is used to statistically analyze the sowing state at different times, obtain a state sequence, perform temporal and feature analysis on the state sequence, and correct the sowing state. The feature analysis includes spatiotemporal analysis to obtain instantaneous features and persistent features. Instantaneous features are used to characterize the seed state in the image, and persistent features represent the changes in the seed state over a period of time. When extracting persistent features, the current image is used as the tail image to obtain an image sequence over a period of time, which is used to perform the persistent feature extraction process. The information statistics and feedback module is used to statistically analyze the corrected sowing status of each seeding hole, combine it with images to construct visual information, and feed it back to the main end of the seeder; the visual information includes at least images under different sowing conditions and statistical information on the status of each seeding hole.

[0014] As a further aspect of the present invention: the image acquisition module includes: The signal acquisition unit is used to acquire the seed passage signal output by the seeder alarm monitor; the seed passage signal is used to indicate whether there is an object passing through the seed discharge hole; The instruction sending unit is used to generate a shooting instruction based on the seed signal and send the shooting instruction to the industrial camera device built into the seed metering device of the seeder; The time stamping unit is used to timestamp the images captured by industrial camera equipment in the seed dispensing holes during the rotation of the seed tray. The proprietary frequency determination unit is used to count the time points when the seed passes through the signal in real time and determine the proprietary frequency of the industrial camera equipment; the proprietary frequency is used to acquire images at regular intervals when there is no shooting instruction.

[0015] As a further aspect of the present invention: the sowing status determination module includes: The preprocessing unit is used to read the latest image from any seeding hole in real time and preprocess the image; the preprocessing includes image cropping, size normalization and noise suppression; The region segmentation unit is used to locate the seed region in the preprocessed image, segment the seed region based on the image segmentation algorithm, and extract the seed target and its contour features in the seeding hole. The determination execution unit is used to determine the sowing status based on the extracted seed target and its contour features.

[0016] As a further aspect of the present invention: the sowing state correction module includes: The state statistics unit is used to count the sowing status at different times to obtain a state sequence; The state transition unit is used to perform numerical transformation on the state sequence to obtain a numerical sequence; An abnormal moment marking unit is used to calculate the difference sequence of a numerical sequence and determine the marking of abnormal moments based on the difference sequence; the abnormal moment refers to the moment point corresponding to the element index when the absolute value of the difference reaches a preset threshold. The instantaneous feature determination unit is used to read the seed targets and their contour features of the image at an abnormal moment, and determine the geometric features based on the seed targets and their contour features as instantaneous features; wherein, the instantaneous features include the number of seeds, the proportion of holes occupied, and the seed spacing; The persistent feature determination unit is used to determine an image sequence with the latest image as the tail image based on a preset time span, perform differential recognition on the image sequence, and determine the seed change vector as a persistent feature; The correction execution unit is used to determine the current sowing state of the seed dispensing hole based on the extracted instantaneous and continuous features, and use this as the corrected sowing state; the sowing state includes at least the normal seed suction state, the double seed suction state, the no seed suction state, and the foreign object blockage state.

[0017] Compared with existing technologies, the advantages of this invention are as follows: This invention uses a combination of camera monitoring and alarm monitoring. The driver only needs to observe the display screen to see the seed suction status on the seed tray of the seed metering device through the camera. For example, whether double seeds are being sucked, whether no seeds are being sucked, or whether foreign objects are blocking the small holes of the seed tray. Even if there is dust inside the seed guide tube during operation, causing the alarm monitoring device to overcount the number of seeds, farmers can distinguish between false alarms and normal counting by observing the display screen. This greatly reduces the yield loss for farmers and also solves farmers' doubts about the inaccuracy of the alarm monitoring device. In addition, the entire process has very low requirements for the camera and recognition algorithm. It adopts a gradient recognition architecture, which has high accuracy while consuming very few resources. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention.

[0019] Figure 1 This is a flowchart of a multi-dimensional seeding alarm method for seeders.

[0020] Figure 2 This is a structural diagram of a multi-dimensional seeding alarm system used in seeders. Detailed Implementation

[0021] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0022] Figure 1 This is a flowchart of a multi-dimensional seeding alarm method for a seeder. In this embodiment of the invention, a multi-dimensional seeding alarm method for a seeder includes: Step S100: Images of each seeding hole during the rotation of the seed tray are captured using an industrial camera device built into the seed metering device of the seeder; wherein the image acquisition frequency of each seeding hole is different; The seeder's seed metering device is equipped with an industrial camera. The camera is used to acquire images from the seed holes as the seed disc rotates. The image acquisition frequency of the industrial camera is different for each seed hole. This means that the industrial camera needs to record a time stamp when acquiring images; otherwise, the time information of the image cannot be known due to the variable frequency.

[0023] Step S200: Perform target recognition on the image of each seeding hole, locate the seed target and its outline features, and determine the sowing status; For the images of each seeding hole, the images are identified, and the seeds (image contours) in the images are located. Contour features are synchronized, and the location results and contour features are analyzed to determine the sowing status. It is worth mentioning that locating seed targets and contour recognition are actually similar image processing processes. If the contour features are extracted, it is also a feasible technical solution to determine whether it is a seed. The difference is that the location process may be a low-precision bounding box process. That is, the object in the image is first bounded, called the seed target, and then the contour of the seed target is recognized to determine the sowing status.

[0024] Step S300: Statistically analyze the sowing status at different times to obtain a status sequence, perform temporal and feature analysis on the status sequence, and correct the sowing status; the feature analysis includes spatiotemporal analysis to obtain instantaneous features and persistent features. Instantaneous features are used to characterize the seed status in the image, and persistent features represent the changes in the seed status over a period of time. When extracting persistent features, the current image is used as the tail image to obtain an image sequence over a period of time, which is used to perform the persistent feature extraction process. The image itself contains a time stamp. Image recognition yields the sowing status at the corresponding time. Statistical analysis of the sowing status at different times creates a state sequence, representing the seed passage through a sowing hole at different times. However, due to potential issues in the recognition process itself, temporal and feature analysis of the state sequence is necessary to correct the sowing status. This temporal and feature analysis process means that the accuracy of sowing status recognition does not need to be excessively high. In reality, seeds pass through the seed discharge holes at a very high speed. If only sowing status recognition is used, extremely high precision and computing power are required to accurately identify clear images in a very short time. This is clearly unsuitable for a seeder. The seeder's recognition process is merely an auxiliary function; what is needed is... Minimal resources are required; the introduction of temporal and feature analysis processes can significantly reduce the accuracy requirements for sowing status recognition. Even with a high error rate (in current technologies, image recognition accuracy is already high, such as over 90%. Reducing accuracy from 99% to 90% is a huge reduction; in other words, increasing accuracy from 90% to 99% requires significantly higher costs), the subsequent feature analysis process can correct the error. This feature analysis process is essentially a data sequence analysis, requiring only numerical processing and consuming minimal computational power. Under this architecture, the entire visual recognition process for sowing requires minimal resources and places very low demands on the performance of industrial cameras and processors.

[0025] Step S400: Statistically analyze the corrected sowing status of each seeding hole, construct visualization information by combining the image, and feed it back to the main end of the seeder; wherein, the visualization information includes at least images under different sowing statuses and statistical information on the status of each seeding hole; The corrected sowing status of each seed hole is the final sowing status. Combined with the acquired images, the sowing status and images are converted into visual information and fed back to the seeder's central terminal, which is generally located in the cab. After being converted into visual information, it can be quickly displayed using the display device in the cab, allowing the operator to understand the sowing situation in real time.

[0026] Regarding step S100, the step of acquiring images of each seed dispensing hole during the rotation of the seed disc based on an industrial camera device built into the seed metering device of the seeder includes: Collect the seed passage signal output by the alarm monitor of the seeder; the seed passage signal is used to indicate whether there is an object passing through the seed discharge hole; Based on the seed, a shooting command is generated through a signal and sent to an industrial camera device built into the seed metering device of the seeder; Images of the seed discharge holes during the rotation of the seed tray are captured using industrial camera equipment, and the captured images are timestamped. The system calculates the time points at which the seed passes through the signal in real time to determine the proprietary frequency of the industrial camera equipment; the proprietary frequency is used to periodically acquire images when there is no shooting instruction.

[0027] In one embodiment of the technical solution of this invention, the image acquisition process is specifically described. The process involves acquiring a seed passage signal output by the seeder alarm monitor. This seed passage signal is a trigger signal, for example, determined by real-time measurement of whether there is an obstruction at the entrance using an infrared sensor, thus indicating whether an object is passing through. This seed passage signal is used to characterize whether an object is passing through the seed metering hole. Each time a seed passage signal is detected, a shooting command is generated and sent to an industrial camera device built into the seeder's seed metering unit. The industrial camera device then acquires an image either once or continuously over a period of time (at a preset maximum frequency). This process provides active image acquisition. In the control process, if no seeds pass through, image acquisition will not occur, which may lead to some missed information. Therefore, the above-mentioned method provides a passive real-time acquisition process by setting an independent frequency for the industrial camera equipment. Even if no shooting command is generated, the industrial camera equipment will still acquire images periodically. As for the independent frequency, it can certainly be a fixed frequency, but a fixed frequency may not be high enough in terms of resource utilization. For example, if images are continuously acquired during a long period of inactivity, resources will still be consumed. Therefore, the technical solution of this invention determines the temporal distribution of seed passage by statistically analyzing the time points of seed passage signals in real time, and then determines an independent frequency that fits the actual situation. This is a dynamic frequency that greatly improves resource utilization, allowing more resources to be used during more frequent sowing periods.

[0028] Regarding step S200, the step of performing target recognition on the image of each seeding hole, locating the seed target and its contour features, and determining the sowing status includes: For any seeding hole, the latest image is read in real time, and the image is preprocessed; the preprocessing includes image cropping, size normalization, and noise suppression; In the preprocessed image, the seed region is located, and the seed region is segmented based on the image segmentation algorithm to extract the seed target and its contour features in the seeding hole; The sowing status is determined based on the extracted seed targets and their contour features.

[0029] In one example of the technical solution of this invention, the process of determining the sowing status is specifically described. For any seeding hole, the latest image is read in real time, and the image is preprocessed. The preprocessing includes image cropping, size normalization, and noise suppression, which are conventional pre-image processing operations and will not be described in detail here. The seed region is located in the preprocessed image, and the seed region is segmented based on an image segmentation algorithm. The seed targets and their contour features within the seeding holes are extracted. Extracting the seed targets is a box selection process, and extracting the contour features is a process of recognizing the contour within the box. The sowing status can be determined based on the extracted seed targets and their contour features.

[0030] Regarding step S300, the steps of statistically analyzing the sowing status at different times to obtain a state sequence, performing time-domain analysis and feature analysis on the state sequence, and correcting the sowing status include: The sowing status at different times is statistically analyzed to obtain a state sequence; A numerical sequence is obtained by performing a numerical transformation on the state sequence; Calculate the difference sequence of the numerical sequence, and determine the marked abnormal time based on the difference sequence; the abnormal time refers to the time point corresponding to the element index when the absolute value of the difference reaches a preset threshold. The seed targets and their contour features of the image at the abnormal moment are read, and the geometric features are determined based on the seed targets and their contour features as instantaneous features; wherein, the instantaneous features include the number of seeds, the proportion of holes occupied, and the seed spacing; Based on a preset time span, an image sequence is determined with the image at the latest moment as the tail image. The image sequence is then subjected to differential recognition to determine the seed change vector, which serves as a persistent feature. Based on the extracted instantaneous and continuous features, the current sowing state of the seed discharge hole is determined as the corrected sowing state; the sowing state includes at least the normal seed suction state, the double seed suction state, the no seed suction state, and the foreign object blockage state.

[0031] The above describes the correction process for the sowing state. Sowing states at different times are statistically analyzed to obtain a state sequence. Since sowing states are generally state labels in text format, they are not suitable for numerical processing. Therefore, the state sequence needs to be converted first, transforming each sowing state into a numerical value, thus obtaining a numerical sequence. For the obtained numerical sequence, the difference sequence is calculated. The difference sequence represents the changes in each value in the numerical sequence, which in practice indicates changes in the sowing state. Analyzing the difference sequence can mark abnormal moments. An abnormal moment refers to the time point corresponding to the element index when the absolute value of the difference reaches a preset threshold, i.e., a sudden change in the sowing state. The seed target and its contour features of the image at the abnormal moment are read, and based on the seed target and its contour... The process of determining geometric features, as instantaneous features, is essentially a more precise image recognition process. Instantaneous features are the features of an image at a single moment, hence the name. Further, using the latest image as the tail image, images within a preset time span are read forward to obtain an image sequence. Differential recognition is performed on the image sequence to determine the change vector of each seed contour (the position vector can be determined based on the position of the contour center point, called the change vector). It is a feature over a period of time, hence the name persistent feature. Based on the extracted instantaneous and persistent features, it is determined whether the current sowing state of the seeding hole is accurate. If the two are different, the sowing state determined based on the extracted instantaneous and persistent features is taken as the final sowing state, that is, as the corrected sowing state.

[0032] Regarding step S400, the step of statistically analyzing the corrected sowing status of each seeding hole, constructing visual information by combining it with the image, and feeding it back to the main end of the seeder includes: Read the state sequence and its difference sequence; The state sequence is segmented according to the difference sequence, and the image display ratio within each time period is determined synchronously. Based on the image display ratio, images are randomly extracted at different time periods, and images at abnormal moments are extracted synchronously. The state sequence and extracted images are used as visualization information and fed back to the central terminal of the seeder.

[0033] In one example of the technical solution of this invention, the construction and feedback process of visualization information is specifically described. The process involves reading the state sequence and its difference sequence, segmenting the state sequence according to the difference sequence, and using the segmentation method as follows: the sequence number in the difference sequence where the absolute value of the difference is greater than a preset threshold is located and used as the segmentation time point. For two adjacent segmentation time points, if there are enough sequence numbers between them, then they can be considered as a time period. The image display ratio is determined based on the number of sequence numbers in each time period, and the image display ratio is directly proportional to the number of sequence numbers. This means that the longer the duration of a certain state, the more images need to be extracted in that time period. This extraction is for regular images. Simultaneously, for images from abnormal times, these are essential feedback contents that must be extracted and uploaded in a separate process.

[0034] As a preferred embodiment of the technical solution of the present invention, the method further includes: Establish a connection channel with the sensors built into the seeder, acquire data from each sensor based on a preset data template, and obtain a real-time matrix; Read the seeding status at each time point and construct a real-time matrix to a sample set of seeding status; The pre-detection model is trained based on the sample set, and the default frequency of the industrial camera is recursively adjusted according to the error rate of the pre-detection model.

[0035] In one embodiment of the technical solution of this invention, a pre-identification scheme is also introduced. The seeder itself has a large number of sensors, which can be used to acquire operational data, reflecting the working status of the seeder. A connection channel is established with the sensors built into the seeder, and data from each sensor is acquired based on a preset data template to obtain a real-time matrix. The core of the data template is the correspondence between the row and column positions and the sensors. One row and column position corresponds to one sensor. Once determined, the data template is not changed. The correspondence method is not complicated; it can be arranged sequentially according to the sensor number, corresponding to each row and column position. The data template acquires data from each sensor to obtain a matrix for each time moment, called the real-time matrix. Specifically, it typically involves first setting time points, then for each row and column position, querying the data from the sensor at that position at the most recent time, and inserting it into the matrix. Simultaneously, the sowing status at the same time is read, constructing a sample set from the real-time matrix to the sowing status. A pre-detection model is trained based on this sample set, and the inherent frequency of the industrial camera is recursively adjusted according to the error rate of the pre-detection model. This adjustment process is recursive; the pre-detection model is a mapping from the seeder's sensor data to the sowing status. This process does not involve image acquisition and recognition. The process simply involves reading data, inputting the mapping relationship, and performing calculations. This is essentially a one-dimensional processing step, consuming far fewer resources compared to image recognition. Therefore, considering the error rate of the pre-detection model, a higher error rate requires more image acquisition, while a lower error rate requires less. The amount of image acquisition used depends on two main processes: active image capture (detecting the seed signal) and passive image capture based on the image's own frequency. The former is not adjustable, while the latter is. The frequency is directly proportional to the error rate, meaning a higher error rate requires more frequent image acquisition to maintain accuracy. The real-time performance of the verification is enhanced; the smaller the error rate, the more the pre-detection model can be applied, reducing the self-frequency and decreasing the number of times high-resource-demand processes are used within the preset time period. Of course, the priority of this process and the self-frequency adjustment process in step S100 is a concern. Generally, these can be considered as two influencing factors to determine the self-frequency simultaneously. Alternatively, a priority can be set, with the self-frequency adjustment process based on the error rate having a higher priority. In this case, the self-frequency is affected by the error rate. The optimal solution is to pre-set weights to adjust both the self-frequency adjustment process based on the time point and the self-frequency adjustment process based on the error rate.

[0036] Figure 2 This is a structural diagram of a multi-dimensional seeding alarm system for a seeder. In this embodiment of the invention, a multi-dimensional seeding alarm system for a seeder, system 10, includes: The image acquisition module 11 is used to acquire images of each seeding hole during the rotation of the seed tray based on an industrial camera device built into the seed metering device of the seeder; wherein, the image acquisition frequency of each seeding hole is different; The sowing status determination module 12 is used to perform target recognition on the image of each seeding hole, locate the seed target and its outline features, and determine the sowing status. The sowing state correction module 13 is used to statistically analyze the sowing state at different times, obtain a state sequence, perform temporal and feature analysis on the state sequence, and correct the sowing state. The feature analysis includes spatiotemporal analysis to obtain instantaneous features and persistent features. Instantaneous features are used to characterize the seed state in the image, and persistent features represent the changes in the seed state over a period of time. When extracting persistent features, the current image is used as the tail image to obtain an image sequence over a period of time, which is used to perform the persistent feature extraction process. The information statistics feedback module 14 is used to statistically analyze the corrected sowing status of each seeding hole, combine the image to construct visualization information, and feed it back to the main end of the seeder; wherein, the visualization information includes at least images under different sowing statuses and status statistics information of each seeding hole.

[0037] Furthermore, the image acquisition module 11 includes: The signal acquisition unit is used to acquire the seed passage signal output by the seeder alarm monitor; the seed passage signal is used to indicate whether there is an object passing through the seed discharge hole; The instruction sending unit is used to generate a shooting instruction based on the seed signal and send the shooting instruction to the industrial camera device built into the seed metering device of the seeder; The time stamping unit is used to timestamp the images captured by industrial camera equipment in the seed dispensing holes during the rotation of the seed tray. The proprietary frequency determination unit is used to count the time points when the seed passes through the signal in real time and determine the proprietary frequency of the industrial camera equipment; the proprietary frequency is used to acquire images at regular intervals when there is no shooting instruction.

[0038] Specifically, the sowing status determination module 12 includes: The preprocessing unit is used to read the latest image from any seeding hole in real time and preprocess the image; the preprocessing includes image cropping, size normalization and noise suppression; The region segmentation unit is used to locate the seed region in the preprocessed image, segment the seed region based on the image segmentation algorithm, and extract the seed target and its contour features in the seeding hole. The determination execution unit is used to determine the sowing status based on the extracted seed target and its contour features.

[0039] Furthermore, the sowing status correction module 13 includes: The state statistics unit is used to count the sowing status at different times to obtain a state sequence; The state transition unit is used to perform numerical transformation on the state sequence to obtain a numerical sequence; An abnormal moment marking unit is used to calculate the difference sequence of a numerical sequence and determine the marking of abnormal moments based on the difference sequence; the abnormal moment refers to the moment point corresponding to the element index when the absolute value of the difference reaches a preset threshold. The instantaneous feature determination unit is used to read the seed targets and their contour features of the image at an abnormal moment, and determine the geometric features based on the seed targets and their contour features as instantaneous features; wherein, the instantaneous features include the number of seeds, the proportion of holes occupied, and the seed spacing; The persistent feature determination unit is used to determine an image sequence with the latest image as the tail image based on a preset time span, perform differential recognition on the image sequence, and determine the seed change vector as a persistent feature; The correction execution unit is used to determine the current sowing state of the seed dispensing hole based on the extracted instantaneous and continuous features, and use this as the corrected sowing state; the sowing state includes at least the normal seed suction state, the double seed suction state, the no seed suction state, and the foreign object blockage state.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-dimensional seeding alarm method for a seeder, characterized in that, The method comprises: Collecting images of each seed hole in the process of seed disc rotation based on an industrial camera device built in the seed metering device of the seeding machine; wherein the image collection frequency of each seed hole is different; Performing target recognition on the images of each seed hole, locating seed targets and their contour features, and determining the seeding state; Statistically analyzing the seeding state at different time points to obtain a state sequence, performing time domain analysis and feature analysis on the state sequence, and correcting the seeding state; the feature analysis comprises space-time analysis to obtain instantaneous features and persistent features, the instantaneous features are used to represent the seed state in the image, and the persistent features represent the change of the seed state within a period of time; in the extraction of the persistent features, the current image is taken as a tail image to obtain an image sequence within a period of time, which is used to perform a persistent feature extraction process; Statistically analyzing the corrected seeding state of each seed hole, combining the images to construct visual information, and feeding back to the total end of the seeding machine; wherein the visual information at least includes images under different seeding states and state statistical information of each seed hole.

2. The multi-dimensional seeding alert method for a seeding machine of claim 1, wherein, The step of collecting images of each seed hole in the process of seed disc rotation based on an industrial camera device built in the seed metering device of the seeding machine comprises: Collecting a seed passing signal output by a seeding machine alarm monitor; the seed passing signal is used to represent whether there is an object passing through the seed hole; Generating a shooting instruction based on the seed passing signal and sending the shooting instruction to the industrial camera device built in the seed metering device of the seeding machine; Collecting images in the seed hole in the process of seed disc rotation based on the industrial camera device, and marking the collected images with time stamps; Real-time statistical analysis of the time point of the seed passing signal to determine the self-frequency of the industrial camera device; the self-frequency is used to collect images at regular intervals when there is no shooting instruction.

3. The multi-dimensional seeding alert method for a seeding machine of claim 1, wherein, The step of performing target recognition on the images of each seed hole, locating seed targets and their contour features, and determining the seeding state comprises: For any seed hole, real-time reading of the image at the latest time, and pre-processing of the image; the pre-processing includes image cropping, size normalization and noise suppression; Locating the seed area in the pre-processed image, segmenting the seed area based on an image segmentation algorithm, and extracting the seed target and its contour features in the seed hole; Determining the seeding state according to the extracted seed target and its contour features.

4. The multi-dimensional seeding alert method for a seeding machine of claim 1, wherein, The step of statistically analyzing the seeding state at different time points to obtain a state sequence, performing time domain analysis and feature analysis on the state sequence, and correcting the seeding state comprises: Statistically analyzing the seeding state at different time points to obtain a state sequence; Performing numerical conversion on the state sequence to obtain a numerical sequence; Calculating the difference sequence of the numerical sequence, and determining the marked abnormal time point according to the difference sequence; the abnormal time point refers to the time point corresponding to the element number when the difference absolute value reaches a preset threshold; Reading the seed target and its contour features of the image at the abnormal time point, determining the geometric features according to the seed target and its contour features as the instantaneous features; wherein the instantaneous features include the number of seeds, the proportion of seeds in the hole and the seed spacing; Determining the image sequence with the latest image as the tail image based on a preset time span, performing difference recognition on the image sequence, determining the seed change vector as the persistent features, and performing difference recognition on the image sequence. Determine the current seeding state of the seed hole as the corrected seeding state based on the extracted instantaneous feature and the persistent feature; the seeding state at least includes a normal seed suction state, a double seed suction state, a non-seed suction state, and a foreign object blockage state.

5. The multi-dimensional seeding alert method for a seeding machine of claim 4, wherein, The step of statistically correcting the seeding state of each seed hole and combining the images to construct visual information and feeding back to the total end of the seeding machine includes: reading the state sequence and the difference sequence; segmenting the state sequence according to the difference sequence, and synchronously determining the image display ratio in each time period; randomly extracting images in different time periods based on the image display ratio, and synchronously extracting images at abnormal moments; feeding back the state sequence and the extracted images to the visual information to the total end of the seeding machine.

6. The multi-dimensional seeding alert method for a seeding machine of claim 1, wherein, The method further includes: establishing a connection channel with the sensors built in the seeding machine, obtaining data of each sensor based on a preset data template, and obtaining a real-time matrix; reading the seeding state at each moment, and constructing a sample set of the real-time matrix to the seeding state; training a pre-detection model based on the sample set, and recursively adjusting the self-frequency of the industrial camera device according to the error rate of the pre-detection model.

7. A multi-dimensional seeding alarm system for a seeder, characterized in that, The system includes: an image acquisition module, configured to acquire images of each seed hole in the seed disc rotation process based on an industrial camera device built in the seed dispenser of the seeding machine; wherein the image acquisition frequency of each seed hole is different; a seeding state determination module, configured to perform target identification on the images of each seed hole, locate the seed target and its contour features, and determine the seeding state; a seeding state correction module, configured to statistically correct the seeding state at different moments to obtain a state sequence, perform time domain analysis and feature analysis on the state sequence, and correct the seeding state; the feature analysis includes space-time analysis to obtain instantaneous features and persistent features; the instantaneous features are used to represent the seed state in the image, and the persistent features represent the change of the seed state in a period of time; in the extraction of the persistent features, the current image is taken as a tail image to obtain an image sequence in a period of time, which is used to perform a persistent feature extraction process; an information statistical feedback module, configured to statistically correct the corrected seeding state of each seed hole, combine the images to construct visual information, and feed back to the total end of the seeding machine; wherein the visual information at least includes images under different seeding states and state statistical information of each seed hole.

8. The multi-dimensional seeding alert system for a planter of claim 7, wherein, The image acquisition module includes: a signal acquisition unit, configured to acquire a seed passing signal output by a seeding machine alarm monitor; the seed passing signal is used to represent whether there is an object passing through the seed hole; an instruction sending unit, configured to generate a shooting instruction based on the seed passing signal, and send the shooting instruction to an industrial camera device built in the seed dispenser of the seeding machine; a time marking unit, configured to acquire images in the seed hole in the seed disc rotation process based on the industrial camera device, and mark the acquired images with time stamps; a self-frequency determination unit, configured to statistically correct the time point of the seed passing signal in real time, and determine the self-frequency of the industrial camera device; the self-frequency is used to acquire images at regular intervals when there is no shooting instruction.

9. The multi-dimensional seeding alert system for a planter of claim 7, wherein, The seeding state determination module includes: The pre-processing unit is configured to read a latest image in real time for any seed sowing hole, and pre-process the image; the pre-processing includes image cropping, size normalization and noise suppression; The region segmentation unit is configured to locate a seed region in the pre-processed image, segment the seed region based on an image segmentation algorithm, and extract a seed target and contour features of the seed target in the seed sowing hole; The determination execution unit is configured to determine a sowing state according to the extracted seed target and contour features.

10. The multi-dimensional seeding alert system for a planter of claim 7, wherein, The sowing state correction module includes: The state statistics unit is configured to count sowing states at different time points to obtain a state sequence; The state conversion unit is configured to perform numerical conversion on the state sequence to obtain a numerical sequence; The abnormal time point marking unit is configured to calculate a difference sequence of the numerical sequence, and determine a marked abnormal time point according to the difference sequence; the abnormal time point refers to a time point corresponding to an element sequence number when a difference absolute value reaches a preset threshold; The instantaneous feature determination unit is configured to read a seed target and contour features of an image at the abnormal time point, and determine geometric features of the seed target and contour features as instantaneous features; the instantaneous features include a seed quantity, a hole proportion and a seed spacing; The continuous feature determination unit is configured to determine an image sequence with the latest image as a tail image based on a preset time span, perform difference identification on the image sequence, and determine a seed change vector as a continuous feature; The correction execution unit is configured to determine a current sowing state of the seed sowing hole based on the extracted instantaneous features and continuous features as a corrected sowing state; the sowing state includes at least a normal seed suction state, a double seed suction state, a non-seed suction state and a foreign matter blockage state.

Images