A corn seed posture estimation method and system

Abstract

The application relates to the technical field of precision seeding, and discloses a corn seed posture estimation method and system. The method comprises the following steps: arranging a white light emitting array and a photoelectric receiving array; each emitting unit is loaded with a modulation signal of different frequencies to output a multi-frequency modulation white light signal; the seed blocks the light path, and the receiving array outputs a composite light intensity response signal; the signal is amplified, filtered and frequency-selected to separate the shielding response signals of each channel; the time sequence shielding response of each channel is structured and organized into a shielding signal feature matrix; time and space synchronization registration is performed on the posture label to construct a supervised data set; an iterative training posture mapping model is constructed to establish a nonlinear mapping relationship between the shielding signal and the three-dimensional posture; real-time signals are collected to input the posture mapping model, and three-dimensional posture parameters are output. The application solves the problems of poor distinguishability and lack of systematic calibration mechanism of the existing photoelectric detection posture, and improves the precision and real-time performance of the three-dimensional posture estimation of high-speed falling corn seeds.

Description

Technical Field

[0001] This application relates to the field of precision seeding technology, and in particular to a method and system for estimating the attitude of maize seeds. Background Technology

[0002] In precision seeding operations, corn seeds undergo random tumbling motion during seeding, seed guiding, and free fall. Their orientation upon entering the soil directly affects the direction of germ growth and the uniformity of seedling emergence. Therefore, real-time detection of the three-dimensional orientation of corn seeds during the fall process is a key technology for achieving directional seeding and closed-loop control of seeding quality.

[0003] Existing machine vision-based posture detection methods typically use high-speed cameras to capture images of falling seeds and then recover posture information through 3D reconstruction algorithms or deep learning networks. However, dust interference and mechanical vibrations at the sowing site make it difficult to guarantee image quality; the high speed of seed falling necessitates high-speed cameras and auxiliary light sources, resulting in high system costs and large data throughput, making real-time processing on embedded platforms difficult; multi-view stereo matching exhibits significant errors when seed texture is scarce, leading to unstable posture calculations; deep learning relies on a large number of manually annotated posture images, resulting in poor annotation consistency and limited model generalization ability. Existing photoelectric sensing-based detection methods use single-optical-path or light curtain arrays to perform threshold decisions based on light intensity changes, achieving presence detection or coarse position estimation. These methods can only acquire one-dimensional information such as projection width or occlusion duration, losing crucial posture information such as seed orientation and embryo surface direction; the same occlusion signal may correspond to multiple different postures, resulting in severely insufficient posture discriminability. Existing multi-channel photoelectric systems mostly use fixed-frequency driving or time-division multiplexing. The former cannot separate the frequency domains of each channel, easily causing crosstalk, while the latter sacrifices temporal resolution, making it difficult to capture continuous posture changes during high-speed falling. In addition, existing methods generally lack a systematic calibration mechanism between sensor signals and the actual pose. The mapping relationship often relies on manual experience or simplified geometric model settings, which makes it difficult to cover the entire pose space where the seed changes continuously, resulting in significant bottlenecks in accuracy and generalization ability. Summary of the Invention

[0004] To address the above shortcomings, this application provides a method and system for estimating the attitude of maize seeds, which solves the problems of poor attitude identification and lack of systematic calibration mechanism in existing photoelectric detection methods, and improves the accuracy and real-time performance of three-dimensional attitude estimation for high-speed falling maize seeds.

[0005] Firstly, this application provides a method for estimating the posture of maize seeds, including: S1. A detection area is formed by arranging a white light emitting array and a photoelectric receiving array opposite each other on both sides of the corn seed falling path. Each white light emitting unit of the white light emitting array constitutes an independent emission channel, and each photoelectric receiving unit of the photoelectric receiving array constitutes an independent receiving channel. S2. The modulation signal generation module generates modulation signals with different frequencies and loads them onto each white light emitting unit, outputting a multi-frequency modulated white light signal with a unique frequency domain identifier. S3. When corn seeds pass through the detection area, they block the multi-frequency modulated white light signals in different detection optical paths. The photoelectric receiving array outputs a composite light intensity response signal, which is then amplified, filtered, and frequency-selected to separate the blocking response signals of each transmission channel. S4. Organize the blocking response signals of each transmitting channel and each receiving channel in a structured manner in the time series to generate a blocking signal feature matrix that can characterize the dynamic blocking changes of the seed. S5. Perform time and space synchronization registration between the occlusion signal feature matrix and the attitude labels generated by the three-dimensional attitude calibration system to construct a signal attitude pairing supervision dataset. S6. Based on the signal attitude pairing supervision dataset, iterate the attitude mapping model to establish a nonlinear mapping relationship between the occlusion signal features and the three-dimensional attitude parameters of the corn seed, and obtain the trained attitude mapping model. S7. During sowing operations, collect real-time occlusion signals, construct a real-time occlusion signal feature matrix, input it into the trained posture mapping model, and output the three-dimensional posture parameters of the corn seeds for sowing control and sensor calibration.

[0006] Secondly, this application provides a maize seed attitude estimation system for implementing the aforementioned maize seed attitude estimation method, comprising: The acquisition and arrangement module is used to arrange a white light emitting array and a photoelectric receiving array opposite each other on both sides of the corn seed falling path to form a detection area. Each white light emitting unit of the white light emitting array constitutes an independent emission channel, and each photoelectric receiving unit of the photoelectric receiving array constitutes an independent receiving channel. The multi-frequency modulation module is used to generate modulation signals with different frequencies through the modulation signal generation module and load them onto each white light emitting unit, outputting a multi-frequency modulated white light signal with a unique frequency domain identifier. The signal separation module is used to block the multi-frequency modulated white light signals in different detection optical paths when corn seeds pass through the detection area. The photoelectric receiving array outputs a composite light intensity response signal, which is amplified, filtered and frequency-selected to separate the blocking response signals of each transmission channel. The signal integration module is used to structure and organize the occlusion response signals of each transmission channel and each receiving channel in the time series, and generate an occlusion signal feature matrix that can characterize the dynamic occlusion changes of the seed. The data establishment module is used to perform time and space synchronization registration between the occlusion signal feature matrix and the attitude labels generated by the three-dimensional attitude calibration system to construct a signal attitude pairing supervision dataset. The model training module is used to iterate the pose mapping model based on the signal pose pairing supervision dataset, establish a nonlinear mapping relationship between occlusion signal features and three-dimensional pose parameters of corn seeds, and obtain the trained pose mapping model. The sowing control module is used to collect real-time occlusion signals during sowing operations, construct a real-time occlusion signal feature matrix and input it into the trained posture mapping model, and output the three-dimensional posture parameters of the corn seeds for sowing control and sensor calibration.

[0007] Thirdly, this application provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the aforementioned corn seed attitude estimation method.

[0008] The beneficial effects of this application are as follows: This application proposes a method and system for estimating the attitude of maize seeds. Compared with existing technologies, this application achieves multi-channel parallel detection through multi-frequency modulation and combines a supervised learning-driven attitude mapping model, solving the problems of poor attitude discernibility and lack of systematic calibration mechanism in existing photoelectric detection methods. Specifically, this application applies different frequency modulation signals to each transmission channel, enabling the receiver to separate the independent occlusion response of each channel from the composite light intensity response through frequency selective filtering. This avoids channel crosstalk while retaining full-band time resolution, solving the problem that time-division multiplexing cannot capture the continuous attitude changes of seeds during high-speed falling. An occlusion signal feature matrix is ​​constructed with the transmission channel, receiving channel, and sampling time sequence as dimensions, enabling the joint expression of spatial occlusion distribution and temporal evolution. The occlusion signal feature matrix is ​​time-space synchronized with the attitude labels output by the three-dimensional attitude calibration system to construct a supervised dataset, solving the problem of lack of systematic calibration mechanism between signal and attitude. Based on this, a mapping model is trained to establish a nonlinear mapping relationship between occlusion signal and attitude parameters. The attitude parameters can be directly output during the inference stage without relying on manual threshold setting or geometric model inversion. The above steps work together to form a complete technical chain from signal acquisition, feature representation, dataset construction to attitude mapping, realizing a systematic method for directly inferring the three-dimensional attitude of maize seeds from photoelectric array signals. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating a method for estimating the attitude of maize seeds according to this application; Figure 2 This is a schematic diagram of the multi-frequency modulated white light array detection device in this application; Figure 3 This is a schematic diagram of the structure of a maize seed attitude estimation system according to this application. Detailed Implementation

[0011] This application provides a method and system for estimating the attitude of corn seeds. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0012] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of a corn seed attitude estimation method in this application includes: S1. A detection area is formed by arranging a white light emitting array and a photoelectric receiving array opposite each other on both sides of the corn seed falling path. Each white light emitting unit of the white light emitting array constitutes an independent emission channel, and each photoelectric receiving unit of the photoelectric receiving array constitutes an independent receiving channel.

[0013] In one specific embodiment, performing S1 includes the following steps: A white light emitting array is arranged on the first side of the corn seed falling path. The white light emitting array includes multiple white light emitting units arranged in a horizontal array along the seed falling direction. A photoelectric receiving array is arranged on the second side of the corn seed falling path. The photoelectric receiving array includes multiple photoelectric receiving units that are arranged corresponding to the white light emitting unit. The white light emitting array and the photoelectric receiving array are arranged opposite each other, so that each white light emitting unit and each photoelectric receiving unit form multiple pairs of oppositely arranged emitting and receiving correspondences in space, and a detection area for corn seeds to fall freely is formed between them.

[0014] Specifically, in precision seeding operations, corn seeds are discharged by a seed metering device and fall freely along a seed guide tube. Their fall path is a nearly vertical channel. To obtain the spatial attitude information of the seeds during this fall, a through-beam photoelectric detection structure needs to be constructed on both sides of the fall path. In practice, a white light emitting array is arranged on the first side of the corn seed's fall path. The white light emitting array is an array-type light source structure composed of multiple white light emitting units. These units can be white light-emitting diodes or broadband light sources. Multiple white light emitting units are arranged along the seed's fall direction to form a one-dimensional linear array, or arranged horizontally perpendicular to the fall direction to form a horizontal linear array. Alternatively, a two-dimensional matrix array or a staggered array can be used. For example, the white light emitting array includes eight white light emitting units, with a center-to-center spacing of 5 mm between each unit. This spacing is smaller than the typical size of a corn seed along its long axis (approximately 10 to 12 mm), ensuring that seeds falling in any posture simultaneously block at least two detection light paths, thus avoiding missed detections. Each white light emitting unit corresponds to an independent emitting channel. That is, each white light emitting unit is driven by its own independent driving circuit, and there is no multiplexing relationship in electrical connection between them. This allows each emitting channel to be independently loaded with a modulation signal of different frequencies, forming a modulated white light output with a unique frequency domain identifier.

[0015] Correspondingly, a photoelectric receiving array is arranged on the second side of the corn seed's falling path, opposite to the first side. The photoelectric receiving array is an array-type receiving structure composed of multiple photoelectric receiving units. These units can be photodiodes, phototransistors, or photosensors, and their function is to convert changes in incident light intensity into electrical signals in the form of current or voltage. Each photoelectric receiving unit corresponds to a white light emitting unit; that is, the installation position of each photoelectric receiving unit is spatially aligned with the installation position of each white light emitting unit. For example, the photoelectric receiving array also includes eight photoelectric receiving units, with a unit spacing of 5 mm. Furthermore, the normal direction of the photosensitive surface of each photoelectric receiving unit coincides with the normal direction of the light-emitting surface of the white light emitting unit, ensuring that the light emitted by the emitting unit travels in a straight line and directly enters the corresponding receiving unit. The relative spacing between the white light emitting array and the photoelectric receiving array is set according to the size of the corn seed and the space through which it falls. For example, it is set to 30 mm. This spacing is larger than the maximum size of the corn seed to ensure that the seed does not come into contact with the arrays on both sides when it falls freely in the detection area, while being small enough to maintain the light intensity signal-to-noise ratio at the receiving end.

[0016] Since the white light emitting array and the photoelectric receiving array are arranged opposite each other, multiple pairs of emission and reception correspondences are formed in space between each white light emitting unit and each photoelectric receiving unit. In addition to the main optical path facing each other, each receiving unit will also receive oblique light from the adjacent emitting unit. Therefore, if the number of emission channels is M and the number of reception channels is N, a total of M×N detection optical paths are formed between the two arrays. These optical paths are distributed crosswise between the emitting side and the receiving side, and the space they enclose together constitutes the detection area, that is, the light field area that the corn seed passes through when it falls freely. When a corn seed enters the detection area, its flattened ellipsoidal or wedge-shaped body blocks the light path passing through its location. The light intensity of the receiving unit corresponding to the blocked light path decreases, while the receiving output of the unblocked light path remains unchanged. Furthermore, the combination of blocked light paths and the blocking depth of each light path are different when the seed is in different postures. For example, when the seed falls vertically with its long axis, it only blocks a few light paths in the middle and the blocking time is short. However, when the seed falls horizontally with its long axis, it blocks multiple horizontally distributed light paths at the same time and the blocking time is long. Therefore, the combination of blocking states of M×N light paths in the detection area carries information about the spatial posture of the seed.

[0017] The array's layout parameters need to match the seed's falling speed. For example, the falling speed of corn seeds at the end of the seed guide tube is approximately 2 to 4 meters per second. The coverage height of the detection area along the falling direction is determined by the array's span along the falling direction. When using an 8-unit linear array with a 5-millimeter spacing, the coverage height is approximately 35 millimeters. The time for a seed to pass through the detection area is approximately 10 to 20 milliseconds. To ensure sufficient sampling points are obtained within this time window, the subsequent sampling frequency is set to above 100 kHz, so that at least 1000 sampling moments are obtained for each seed during its passage, thereby supporting the construction of the temporal dimension of the shading signal feature matrix. Furthermore, to reduce the impact of ambient light and sunlight on the detection, the white light emitting array and photoelectric receiving array can be installed inside a closed detection chamber. The inner wall of the chamber is coated with a light-absorbing layer, ensuring that the light intensity variation output by the receiving unit mainly originates from seed shading rather than ambient light fluctuations.

[0018] S2. The modulation signal generation module generates modulation signals with different frequencies and loads them onto each white light emitting unit, outputting a multi-frequency modulated white light signal with a unique frequency domain identifier.

[0019] In one specific embodiment, performing S2 includes the following steps: The modulation signal generation module generates multiple periodic modulation signals, and the modulation frequencies of each transmission channel are different and meet the preset frequency interval constraints. The signal shaping and filtering module performs waveform shaping, amplitude adjustment and DC bias superposition on the modulation signal to keep the white light emitting unit in a stable light-emitting state. The carrier modulation module loads the processed modulation signal onto a high-frequency carrier signal to form a composite modulation signal, thereby achieving spectrum shifting. The white light driving module controls the driving current of the white light emitting unit according to the composite modulation signal, and outputs a multi-frequency modulated white light signal.

[0020] Specifically, the modulation signal generation module uses direct digital frequency synthesis (DFD) technology to generate multiple periodic modulation signals. This module internally includes a phase accumulator, a waveform storage digital-to-analog converter (DAC), and a low-pass filter. The phase accumulator accumulates the frequency control word under clock drive. The accumulated result is used as an address index to read the corresponding waveform amplitude data from the waveform storage DAC. After being converted into an analog signal by the DAC, the low-pass filter removes clock noise. The modulation signal expression for the m-th transmission channel is as follows: ,in This represents the modulation amplitude of the m-th transmission channel. Indicates the modulation frequency. This represents the initial phase. The modulation signal generation module generates M modulation signals for the M transmission channels in the white light emission array. The modulation frequencies of each transmission channel are different and satisfy a preset frequency spacing constraint. The preset frequency spacing constraint requires that the difference between the modulation frequencies of any two transmission channels satisfies... ,in For the minimum frequency interval, The passband bandwidth of each bandpass filter in the subsequent frequency selection filtering module is determined based on the frequency spacing between adjacent channels to ensure effective separation of different channels. Taking M=8 as an example, the modulation frequencies of the 8 transmit channels are set to 10 kHz, 14 kHz, 18 kHz, 22 kHz, 26 kHz, 30 kHz, 34 kHz, and 38 kHz, respectively, with a frequency spacing of 4 kHz between adjacent channels. The passband bandwidth of each bandpass filter is set in conjunction with this frequency spacing and the passband and stopband performance requirements to avoid spectral aliasing of the signals in the frequency domain.

[0021] The modulated signal output from the modulation signal generation module is sent to the signal shaping and filtering module for processing. The signal shaping and filtering module performs waveform shaping, amplitude adjustment, and DC bias superposition on the modulated signal. Waveform shaping uses a voltage comparator composed of an operational amplifier to convert the sinusoidal modulated signal into a square wave signal of the same frequency; subsequently, band-limited filtering suppresses higher-order odd harmonics in the square wave, retaining the fundamental wave as the effective component for carrier modulation. Amplitude adjustment is achieved using a variable gain amplifier composed of a digital potentiometer and an operational amplifier. Based on the rated drive current range of the white light emitting unit, the peak-to-peak amplitude of the modulated signal is adjusted to a preset amplitude range. This preset amplitude range is determined according to the dynamic drive current range recommended in the white light emitting unit's datasheet, for example, the peak-to-peak amplitude is set between 1 volt and 3 volts. DC bias superposition adds the amplitude-adjusted modulated signal and the DC bias voltage using an adder circuit. The expression for the superimposed signal is: ,in This is the DC bias voltage. The value of the DC bias voltage is set based on the threshold turn-on voltage of the white light emitting unit. For example, if the threshold turn-on voltage of the white light emitting unit is 2.8 volts, then the DC bias voltage is set to 2.8 volts to ensure that the valley voltage of the modulation signal is not lower than the threshold turn-on voltage, so that the white light emitting unit remains in the conducting state throughout the entire modulation cycle without any off-state intervals. The signal after the DC bias is superimposed is amplified by the power amplifier stage circuit. The power amplifier stage circuit adopts a Class AB power amplifier structure, and its quiescent operating point is set at the center of the amplification region to reduce the impact of crossover distortion on the modulation signal waveform.

[0022] The modulated signal processed by the signal shaping and filtering module is sent to the carrier modulation module. The carrier modulation module loads the processed modulated signal onto a high-frequency carrier signal to form a composite modulated signal, achieving spectrum shifting. The carrier modulation module internally includes a high-frequency carrier signal generator and an analog multiplier. The high-frequency carrier signal generator uses a crystal oscillator in conjunction with a phase-locked loop frequency synthesizer to generate the high-frequency carrier signal. ,in The carrier frequency is determined based on the frequency response characteristics of the white light emitting unit. As a light-emitting diode (LED) device, the electro-optical conversion frequency response of the white light emitting unit is relatively flat in the low-frequency range and attenuates with increasing frequency in the high-frequency range. The carrier frequency is set at the boundary between the flat and attenuation regions of the white light emitting unit's frequency response curve, for example, 1 MHz. The analog multiplier receives the processed modulated signal from the signal shaping and filtering module. and high-frequency carrier signal from high-frequency carrier signal generator Multiplying the two together yields the composite modulation signal for the m-th transmission channel. To illustrate the location of the target sideband, the following equivalent analysis is performed on the fundamental component of the square wave after band-limited filtering. Substituting and expanding the fundamental equivalent components, Considering only the DC component and the fundamental square wave component, the expansion shows that the composite modulated signal contains three main frequency components: frequency 1. The carrier component (generated by multiplying the DC bias with the carrier) has a frequency of The upper sideband components and frequencies are The lower sideband component (generated by multiplying the fundamental component of the modulated signal with the carrier). The remaining higher harmonic components after band-limited filtering are not used as channel identification frequencies and are further suppressed by frequency-selective filtering at the receiver. The carrier modulation module shifts the spectrum of the modulated signal from the baseband to the carrier frequency using multiplication operations. The high-frequency band centered on the center allows for the modulation frequencies of different transmission channels. In the frequency domain, they correspond to different upper sideband frequencies. and lower sideband frequency For example, the modulation frequency of the first transmission channel. 10 kHz, carrier frequency If the frequency is 1 MHz, then the upper sideband frequency of this channel is 1010 kHz and the lower sideband frequency is 990 kHz; the modulation frequency of the second transmission channel... If the frequency is 14 kHz, then the upper sideband frequency of this channel is 1014 kHz and the lower sideband frequency is 986 kHz. The sideband frequencies of each transmit channel maintain the same frequency spacing as the modulation frequency in the frequency domain, allowing for signal separation at the receiver through frequency selective filtering. The composite modulation signal output from the carrier modulation module is impedance-matched by a buffer amplifier before being sent to the white light drive module.

[0023] The white light driver module controls the drive current of the white light emitting unit based on the composite modulation signal, outputting a multi-frequency modulated white light signal. The white light driver module includes a voltage-controlled current source circuit and an overcurrent protection circuit. The voltage-controlled current source circuit consists of an operational amplifier, a power transistor, and a sampling resistor forming a negative feedback closed-loop structure. Composite modulation signal The voltage control signal is input to the non-inverting input of the operational amplifier. The output of the operational amplifier drives the base of the power transistor. The emitter of the power transistor is connected in series with a sampling resistor and then grounded. The voltage across the sampling resistor is fed to the inverting input of the operational amplifier via a feedback network. The negative feedback closed-loop structure makes the voltage across the sampling resistor follow the composite modulation signal. The change in current causes a drive current to flow through the sampling resistor and the collector of the power transistor. ,in This refers to the resistance value of the sampling resistor. The resistance value is determined jointly by the rated drive current range of the white light emitting unit and the voltage amplitude range of the composite modulation signal, for example, the composite modulation signal... The voltage amplitude range is 0 volts to 3.3 volts, and the rated drive current range of the white light emitting unit is 0 mA to 100 mA. Therefore, the sampling resistor... Set to 33 ohms to reduce drive current. The range is 0 mA to 100 mA. The overcurrent protection circuit triggers protection via a comparator when the drive current exceeds the rated maximum value, pulling down the base voltage of the power transistor to limit the drive current. The white light emitting unit receives the drive current. Then, its instantaneous output light power With drive current The relationship between them is , where η is the electro-optic conversion efficiency of the white light emitting unit. Due to the driving current ,and Includes DC bias Therefore, the optical power output by the white light emitting unit contains a DC component and an AC modulation component. The DC component keeps the white light emitting unit in a continuous emitting state, while the AC modulation component causes the optical power to vary with the composite modulation signal. The frequency fluctuates according to the changes, thus outputting a multi-frequency modulated white light signal. Each of the M white light emitting units in the white light emitting array independently executes the above driving process. The multi-frequency modulated white light signal output by each white light emitting unit has a unique identifier in the frequency domain, specifically the upper sideband frequency of each channel. and lower sideband frequency The photoelectric receiving array at the receiving end receives a composite optical signal, which is the superposition of M multi-frequency modulated white light signals from the transmitting channels after propagation in space. Each channel signal occupies a different sideband frequency position in the frequency domain. By using a bandpass filter with the center frequency set to the upper or lower sideband frequency of each channel, the independent blocking response signal of each transmitting channel can be separated from the composite optical signal.

[0024] S3. When corn seeds pass through the detection area, they block the multi-frequency modulated white light signals in different detection optical paths. The photoelectric receiving array outputs a composite light intensity response signal, which is then amplified, filtered, and frequency-selected to separate the blocking response signals of each transmission channel.

[0025] In one specific embodiment, performing S3 includes the following steps: When corn seeds enter the detection area between the white light emitting array and the photoelectric receiving array, they simultaneously block multiple spatial detection optical paths between each emitting unit of the white light emitting array and each receiving unit of the photoelectric receiving array. Each photoelectric receiving unit receives the multi-frequency modulated white light signal that arrives after being blocked by corn seeds, converts the optical signal into an electrical signal, and outputs a composite light intensity response signal containing frequency domain components of multiple transmission channels; The composite light intensity response signal is amplified by the preamplifier module and then sent to the frequency selective filter module. The modulation frequency components corresponding to each transmission channel are extracted by multiple bandpass filters with different center frequencies, and the blocking response signal of each transmission channel is obtained.

[0026] Specifically, after being discharged by the seed metering device, the corn seeds enter the detection area between the white light emitting array and the photoelectric receiving array along the seed guide tube. Upon entering the detection area, the seed simultaneously blocks multiple spatial detection optical paths between the emitting units of the white light emitting array and the receiving units of the photoelectric receiving array. Since the white light emitting array contains M emitting units and the photoelectric receiving array contains N receiving units, forming M×N spatial detection optical paths, and the corn seed, being an irregularly flattened ellipsoid or wedge-shaped entity, has different projected contours along its long axis, short axis, and thickness direction under different spatial orientations, the degree of blocking of the optical paths varies at different spatial locations within the detection area. Specifically, when the seed's long axis falls vertically, the occlusion time of the central optical path in the detection area is short and the occlusion width is narrow. When the seed's long axis falls horizontally, it simultaneously occludes multiple horizontally distributed optical paths, and the occlusion width increases. When the seed falls at an angle, the order and duration of occlusion of optical paths at different heights exhibit a differentiated distribution. This differentiated occlusion means that the change in transmitted light intensity corresponding to each of the M×N detection optical paths carries specific information about the seed's spatial attitude. Each white light emitting unit continuously outputs a multi-frequency modulated white light signal with a unique frequency domain identifier. This signal propagates along the straight optical path between each emitting unit and each receiving unit. When the seed does not obstruct a certain optical path, the receiving unit corresponding to that optical path receives the complete modulated white light signal; when the seed obstructs a certain optical path, the seed body absorbs and scatters some of the light in the optical path, resulting in a decrease in the light power reaching the receiving unit.

[0027] Each photoelectric receiving unit receives the multi-frequency modulated white light signal arriving after being blocked by corn seeds, converts the optical signal into an electrical signal, and outputs a composite light intensity response signal containing frequency domain components of multiple transmission channels. The photoelectric receiving unit uses a photodiode, which converts incident light power into photocurrent. The photocurrent has a linear relationship with the incident light power; within a given wavelength range, the photocurrent... ,in The responsivity of the photodiode at wavelength λ. The incident light power reaching the photosensitive surface. Since the multi-frequency modulated white light signals output from each emission channel occupy different sideband frequency positions in the frequency domain, the light power incident on the photodetector unit is the spatial superposition of the light powers from the M emission channels. The photocurrent output by the photodiode contains AC components corresponding to the sideband frequencies of each emission channel and DC components corresponding to ambient light. The N photodetector units in the photodetector array independently complete the above photoelectric conversion process, and the photocurrent signal output by the nth receiver unit... This can be expressed as the current response corresponding to the sum of the optical powers of the M transmitting channels after propagating through their respective optical paths and reaching the receiving unit, i.e. ,in Let be the photoelectric responsivity of the nth receiving channel at wavelength λ corresponding to the mth transmitting channel, with dimensions in A / W. The time-varying function representing the optical power of the multi-frequency modulated white light signal output from the m-th transmission channel: ; Let be the transmission coefficient of the optical path between the m-th transmitting channel and the n-th receiving channel. This transmission coefficient decreases when the seed blocks the path. The photocurrent is caused by ambient light. The N photocurrent signals output by the N photoelectric receiving units constitute N composite light intensity response signals. Each composite light intensity response signal simultaneously contains frequency domain components of M transmission channels. The modulation frequency corresponding to the same transmission channel in each signal is the same, but the amplitude is different. The amplitude difference reflects the difference in the degree of light path obstruction at different spatial locations.

[0028] The N composite light intensity response signals are amplified by a preamplifier module. This module contains N parallel preamplifier channels with identical parameters, each corresponding to the output signal of a photodiode. The preamplifier module employs a transimpedance amplification structure to amplify the photocurrent signal output from the photodiode. Converted to a voltage signal, the feedback resistor of the transimpedance amplifier Determine the gain, output voltage The value of the feedback resistor is determined based on the amplitude range of the photocurrent output by the photodiode and the input voltage range of the analog-to-digital converter. For example, the feedback capacitor and feedback resistor of the transimpedance amplifier are connected in parallel to form a low-pass filter, whose pole frequency is set to be more than twice the highest carrier frequency to ensure that the modulation frequency components of all transmission channels do not experience amplitude attenuation during amplification. The voltage signal output from the preamplifier module is sent to the frequency-selective filter module after passing through a coupling capacitor. The capacitance value of the coupling capacitor is determined based on the lowest modulation frequency to ensure that this frequency component can pass through the coupling capacitor without significant attenuation. For example, the cutoff frequency formed by the coupling capacitor and the input impedance is set to be less than one-tenth of the lowest modulation frequency, allowing all modulation frequency components above this cutoff frequency to pass through. The gain amplification processing of the preamplifier module simultaneously applies to all frequency components in each composite light intensity response signal, including the modulation frequency components of each transmission channel and the DC and low-frequency noise components corresponding to ambient light.

[0029] The composite light intensity response signal, after gain amplification, is sent to a frequency-selective filtering module. This module uses multiple bandpass filters with different center frequencies to extract the modulation frequency components corresponding to each transmission channel, thus separating the blocking response signal for each channel. The frequency-selective filtering module contains a bandpass filter bank for each transmission channel, with the number of bandpass filters equal to the number of transmission channels M. The center frequency of each bandpass filter is set to the sideband frequency of the corresponding transmission channel. Since each transmission channel forms an upper sideband frequency after carrier modulation... and lower sideband frequency With two sideband components, the frequency-selective filtering module can choose to extract one set of sideband frequencies. For example, the center frequencies of M bandpass filters can be uniformly set as the upper sideband frequencies of each channel. Alternatively, the lower sideband frequency of each channel can be uniformly set. The passband bandwidth of the bandpass filter is determined based on the frequency spacing of adjacent channels to effectively separate different channels while ensuring the effective passage of signals from the corresponding channels. The bandpass filter is implemented using an active bandpass filter or a switched-capacitor filter. The order of the bandpass filter is set to 4th order to provide a stopband attenuation slope of 24 dB per octave, effectively suppressing crosstalk between adjacent channel signals. M bandpass filters operate in parallel, each extracting the corresponding frequency component from each composite optical intensity response signal. Since the frequency selective filtering module is connected to the output signals of N receiving channels, it needs to extract M frequency components from each of the N composite signals. Therefore, the frequency selective filtering module contains N×M bandpass filters, each corresponding to a combination of one receiving channel and one transmitting channel. The composite signal of the nth receiving channel passes through a channel with a center frequency of... After the bandpass filter, only the sideband frequency component corresponding to the m-th transmit channel is retained in the output signal. The sideband frequency components of other transmit channels are significantly attenuated by the bandpass filter. The amplitude of the output signal reflects the degree of occlusion of the optical path between the m-th transmit channel and the n-th receive channel by the seed. The frequency selective filtering module outputs M×N signals, where each signal has a specific transmit channel and receive channel identifier. The amplitude of each signal changes dynamically with the seed occlusion state, forming a complete set of occlusion response signals.

[0030] S4. Organize the occlusion response signals of each transmitting channel and each receiving channel in a structured manner on a time series to generate an occlusion signal feature matrix that can characterize the dynamic occlusion changes of the seed.

[0031] In one specific embodiment, performing S4 includes the following steps: The occlusion response signal is demodulated, and the baseband signal corresponding to the occlusion change of each channel is extracted by coherent operation with the reference carrier signal and low-pass filtering. The effective value of the baseband signal is extracted to obtain the blocking response intensity value between each transmit channel and each receive channel; The occlusion response intensity values ​​of all transmit channels and all receive channels at each sampling time are organized in time sequence to construct an occlusion signal feature matrix.

[0032] Specifically, the signal output by the frequency selective filtering module is the blockage response signal of each separated transmit channel and each received channel combination. However, these signals are still in a high-frequency sideband modulation state, and their amplitude envelope carries the blockage change information caused by seed blockage. The frequency components of the carrier itself are unrelated to the blockage change, so the blockage change information needs to be extracted from the high-frequency carrier; this process is demodulation. According to the carrier modulation expansion of S2, the composite modulation signal of the m-th transmit channel contains frequencies of... The upper sideband components and frequencies are The lower sideband component, the center frequency of the bandpass filter of S3 is set to the upper sideband frequency of the corresponding channel. Therefore, the blocking response signal corresponding to the combination of the m-th transmit channel and the n-th receive channel output by the frequency selective filtering module is expressed as: ,in It is a time-varying amplitude envelope that carries real-time information about the degree to which the seed blocks the light path. It is a high-frequency carrier frequency. It is the modulation frequency of the m-th transmission channel. This refers to the signal phase. Demodulation is achieved through coherent computation with a reference carrier signal, which is generated by the carrier synchronization module at the receiver and its frequency is set to the upper sideband frequency of the corresponding transmission channel. Each frequency corresponds one-to-one with the center frequency of the bandpass filter in S3. Its phase is synchronized with the sideband components of the received signal through a phase-locked loop (PLL) circuit. After the PLL is locked, the phase difference between the reference carrier signal and the blocking response signal approaches zero. The reference carrier signal is represented as... The occlusion response signal is multiplied by the reference carrier signal to obtain the product signal. According to the trigonometric identities The product signal is expanded as The first term in the expansion is the difference frequency component, whose frequency is near zero and directly reflects the amplitude envelope. The change is that the second term is the sum-frequency component, whose frequency is twice the sideband frequency. This is a high-frequency component. The product signal is then fed into a low-pass filter. The cutoff frequency of the low-pass filter needs to be greater than the highest change frequency of the occlusion envelope and much less than the double sideband frequency. The highest change frequency of the occlusion envelope is determined by the shortest occlusion duration of the corn seed passing through a single optical path. For example, the seed falling speed is 2 to 4 meters per second, the shortest occlusion time corresponding to a single optical path is about 1 millisecond, and the upper limit of the envelope change frequency is about 1 kHz. However, the double sideband frequency is higher than 2 MHz. Therefore, the cutoff frequency of the low-pass filter is set to 100 kHz, which preserves the baseband component completely while effectively filtering out the double sideband frequency component and high-frequency noise. The low-pass filter adopts a fourth-order Butterworth active filter structure with a passband gain of 1, a stopband attenuation slope of 24 dB per octave, and a passband ripple of less than 0.5 dB. Because after the phase-locked loop locks... The baseband signal output after low-pass filtering is The instantaneous amplitude of the baseband signal is proportional to the amplitude envelope of the blocking response signal. When the seed does not block the light path When the value is large, the seed blocks the light path. The smaller the value, the more severe the occlusion. The smaller the value, the lower the baseband signal. The curve of change over time directly corresponds to the dynamic change in the degree of occlusion of the optical path during the process of the seed falling through the detection area. The M×N occlusion response signals are demodulated one by one to obtain the M×N baseband signals.

[0033] Baseband signal Although the trend of occlusion changes has been reflected, its instantaneous values ​​are significantly affected by noise and signal fluctuations. Directly using instantaneous sampled values ​​to construct the feature matrix leads to feature instability. Therefore, it is necessary to extract the effective value of the baseband signal to obtain the occlusion response intensity value between each transmit channel and each receive channel at each sampling time. The effective value extraction adopts the root mean square calculation method within a sliding time window for the baseband signal of the combination of the m-th transmit channel and the n-th receive channel. The effective value at sampling time t Defined as ,in Indicates the length of the sliding time window. This is the integration variable, representing a time point within the window. Sliding time window length. The settings need to balance smoothing effect and temporal resolution. If the window is too long, the details of occlusion changes will be over-smoothed; if the window is too short, noise cannot be effectively suppressed. The time for corn seeds to pass through the detection area is about 10 to 20 milliseconds, the sampling frequency is set to 100 kHz, and the sliding time window is... Set to 0.5 milliseconds, corresponding to 50 sampling points, this window length is much smaller than the total time it takes for the seed to traverse the detection area. This ensures that the effective value sequence maintains sufficient resolution on the time axis to capture occlusion details during seed pose changes, while effectively suppressing random fluctuations in the baseband signal. In practical engineering implementation, the above integration operation is discretized into a summation operation on the sampling points within the window. ,in This represents the k-th sampling time, and L represents the number of sampling points within the window. , For the sampling frequency, when For 0.5 milliseconds At 100 kHz, L equals 50. Because this calculation uses a causal window, the actual physical time corresponding to the obtained effective value has a lag of approximately half a window, or 0.25 milliseconds. Therefore, the center time of the window is... This effective value serves as a timestamp to eliminate systematic time deviations during time synchronization registration with the attitude tag in S5. Due to differences in the luminous intensity of each transmitting channel, the photoelectric conversion gain of each receiving channel, and the spatial geometric attenuation of each optical path, the absolute dimensions of the effective values ​​of different channel combinations lack comparability. Therefore, the effective values ​​are normalized to the unobstructed reference effective value corresponding to that channel combination when the seed has not entered the detection area. Use the denominator to calculate the normalized occlusion coefficient. The normalized occlusion coefficient is a dimensionless quantity, ranging from 0 to 1. A larger value indicates a less severe occlusion of the optical path, while a smaller value indicates a more severe occlusion. By performing the above effective value extraction and normalization processing on all combinations of M transmit channels and N receive channels, M×N normalized occlusion coefficient sequences are obtained. The length of each sequence is equal to the number of sampling times T. A one-to-one correspondence is maintained between the M×N baseband signals and the M×N normalized occlusion coefficient sequences.

[0034] The normalized blocking coefficients of all transmit and receive channels at each sampling time are organized in time sequence to construct a blocking signal feature matrix. The blocking signal feature matrix is ​​a three-dimensional data structure with three dimensions: transmit channel index m, receive channel index n, and sampling time index k. For each sampling time... The normalized blocking coefficients of all transmit channels and all receive channels at that moment are arranged into an M-row N-column matrix. The row index of this matrix corresponds to the transmit channel number, and the column index corresponds to the receive channel number. The element in the m-th row and n-th column of the matrix... This represents the normalized blocking coefficient between the m-th transmit channel and the n-th receive channel at that sampling time. The matrix corresponding to the T sampling times... Stacked in chronological order, they form a complete occlusion signal feature matrix. Its dimensions are For example, when the white light emitting array contains 8 emission channels and the photoelectric receiving array contains 8 receiving channels, M equals 8, N equals 8, and the matrix at each sampling time is... The matrix is ​​8 rows and 8 columns, with 64 elements, describing the occlusion distribution of the 64 detection optical paths at a given moment. When the seed falls vertically along its major axis, it only occludes a few optical paths in the central detection area. The elements corresponding to the central transmitting and receiving channels have smaller values, while the elements corresponding to the edge channels have values ​​close to 1. When the seed falls horizontally along its major axis, it simultaneously occludes multiple horizontally distributed optical paths. The elements corresponding to multiple rows and columns in the matrix have smaller values, forming a low-value region. As the seed continues to fall within the detection area, its spatial position and attitude angle change, and the degree of occlusion of each optical path also changes accordingly. The values ​​of the elements in the matrix dynamically evolve between different sampling times. After stacking the matrices from T time points in chronological order, the occlusion signal feature matrix is ​​obtained. The system comprehensively records the spatial occlusion distribution and the continuous trajectory of the occlusion distribution along the time axis at every moment during the entire process of a seed entering and leaving the detection area. The structured organization of this matrix enables a joint representation of the spatial occlusion distribution and the temporal evolution process. The spatial dimension is composed of the transmit channel index and the receive channel index, reflecting the occlusion differences of the optical path at different spatial locations within the detection area. The temporal dimension is composed of the sampling time index, reflecting the dynamic change of the occlusion state as the seed falls. Each time slice and the attitude label of the corresponding timestamp in S5 form a one-to-one data basis.

[0035] S5. Perform time and space synchronization registration between the occlusion signal feature matrix and the attitude labels generated by the 3D attitude calibration system to construct a signal attitude pairing supervision dataset.

[0036] In one specific embodiment, performing S5 includes the following steps: The three-dimensional attitude label of the corn seed at each moment during the process of passing through the detection area is obtained by the three-dimensional attitude calibration system. The three-dimensional attitude calibration system is an external calibration device that is independent of the multi-frequency modulation white light array detection device. The attitude label includes the spatial rotation attitude parameters of the corn seed. Using the triggering events of corn seeds entering and leaving the detection area as the time reference, the sampling times of each sampling moment in the occlusion signal feature matrix and the attitude labels output by the three-dimensional attitude calibration system are synchronously aligned on the time axis, so that the occlusion signal features of each sampling moment and the attitude labels of the same moment form a one-to-one correspondence. The coordinate system of the detection area and the coordinate system of the three-dimensional attitude calibration system are spatially unified and calibrated. The attitude label is transformed to the spatial reference coordinate system of the detection area through coordinate transformation, so that the occlusion signal features and the attitude label are in the same coordinate system. The occlusion signal feature matrix after time and space synchronization registration is paired with the corresponding attitude label. Each corn seed passes through the detection area and corresponds to an occlusion signal feature matrix sample and an attitude label sample. The two constitute a paired sample. All paired samples are aggregated to construct a signal attitude pairing supervised dataset.

[0037] Specifically, obtaining 3D pose labels is a prerequisite for constructing a supervised dataset. The 3D pose calibration system is an external calibration device independent of the multi-frequency modulated white light array detection device. Its function is to provide the true pose of the corn seed at each moment during the data acquisition phase. It is only used during dataset construction and is withdrawn after training, as it is no longer needed during actual sowing. For example, the 3D pose calibration system consists of a binocular high-speed camera and a near-infrared structured light projector. The two high-speed cameras are positioned on either side of the detection area, offset from the installation positions of the white light emitting array and the photoelectric receiving array. The structured light projector projects coded light spots onto the detection area. The binocular high-speed camera synchronously acquires images of the falling seeds at a frame rate of 1000 frames per second. The 3D point cloud of the seed surface is recovered through stereo matching, and then iteratively registered with a pre-established 3D corn seed template to calculate the 3D pose label of the seed at each frame. The pose label consists of two parts: spatial rotation pose parameters and spatial position parameters. The spatial rotation pose parameters are exemplarily represented by quaternions. This indicates that the quaternion satisfies the constraint of a modulus of 1, which avoids the gimbal lock singularity problem of Euler angle representation under specific attitudes. It can also be equivalently converted into a rotation matrix R or Euler angles in the form of pitch, yaw, and roll angles. The spatial position parameter is the coordinate vector d of the seed centroid in the three-dimensional coordinate system, and the attitude label at the t-th sampling time is denoted as... For example, the maximum number of iterations for nearest-point registration is set to 50, and the corresponding point distance threshold is set to 0.2 mm. When the change in the mean square error between two adjacent iterations is less than... Convergence is determined when the area is square millimeters, thus ensuring the stability of the attitude label solution.

[0038] Time synchronization registration uses the trigger events of corn seeds entering and leaving the detection area as a unified time reference. Specifically, a pair of independent through-beam trigger light paths are set at the entrance and exit of the detection area along the falling direction. When a seed enters the detection area, it blocks the entrance trigger light path, generating an entry trigger signal. The corresponding time is recorded as follows: When a seed leaves the detection area, it blocks the exit trigger light path, generating a departure trigger signal. The corresponding moment is recorded as... The entry and exit trigger signals are simultaneously sent to the acquisition card of the multi-frequency modulated white light array detection device and the control unit of the three-dimensional attitude calibration system, enabling both systems to share the same hardware time base and eliminating clock drift caused by independent timing. The sampling frequency of the occlusion signal feature matrix is ​​100 kHz, while the output frequency of the three-dimensional attitude calibration system is 1000 frames per second. Their sampling points on the time axis do not overlap, and the timestamp corresponding to each valid value in S4 is the center time of the sliding window. Therefore, using the timestamp of each sampling moment of the occlusion signal feature matrix as a reference, time interpolation is performed on the attitude label sequence output by the three-dimensional attitude calibration system. Cubic spline interpolation is used as an example, interpolating each component of the quaternion and each component of the centroid coordinate separately. After interpolation, the quaternion is normalized again to meet the unit constraint, ensuring that each sampling moment... Slices of the occlusion signal feature matrix Posture label at the same moment A one-to-one correspondence is established. To ensure synchronization quality, an exemplary synchronization error verification condition is set: the difference between the timestamps of the trigger signal in the two systems must not exceed 0.1 milliseconds. Data collected in the current session that exceeds this threshold is discarded and re-collected.

[0039] Spatial registration transforms the attitude label from the 3D attitude calibration system coordinate system to the spatial reference coordinate system of the detection area, ensuring that the occlusion signal features and the attitude label are in the same coordinate system. Specifically, the detection area coordinate system is established with the geometric center of the white light emission array as the origin, the seed falling direction as the vertical axis, and the direction from the transmitting side to the receiving side as the horizontal axis; the 3D attitude calibration system coordinate system is established with the optical center of one of the high-speed cameras as the origin, and there is a fixed rotation and translation relationship between the two. Spatial unified calibration is achieved by placing calibration blocks of known geometric dimensions within the detection area. For example, multiple marker points with known spatial coordinates are distributed on the surface of the calibration block. The 3D attitude calibration system measures the coordinates of each marker point in its coordinate system and pairs them one-to-one with the known coordinates of the marker points in the detection area coordinate system. The coordinate transformation matrix W between the two coordinate systems is solved by least squares. W is a 4×4 homogeneous transformation matrix containing rotation and translation components. The number of marker points is no less than 4 and they are not all coplanar. The reprojection error threshold is set to less than 0.1 mm. If the threshold is met, the calibration result is accepted. After obtaining the coordinate transformation matrix W, the pose labels of each frame output by the 3D pose calibration system are transformed, and the spatial position parameters are adjusted according to... Transform to the detection area coordinate system, and the rotation matrix corresponding to the spatial rotation attitude parameters is as follows: Transformation, where The rotation component in the coordinate transformation matrix W is then converted back to quaternion form and normalized to obtain the attitude label in the detection region coordinate system. Since the coordinate transformation matrix W is constant after the two systems are installed and fixed, spatial registration can be performed again after the device is installed or its position is changed.

[0040] In the pairing and dataset construction phase, the occlusion signal feature matrix, after temporal and spatial synchronization registration, is paired with the corresponding pose label. Each corn seed corresponds to an occlusion signal feature matrix sample throughout its passage through the detection area. and a pose label sample The two constitute a paired sample, where The samples are numbered, and the pose label samples are aligned with the occlusion signal feature matrix samples along the time dimension, and the matrix is ​​sliced. As input to the subsequent pose mapping model, pose labels This serves as the corresponding supervision target. All paired samples are aggregated to construct a signal-pose paired supervision dataset. Where K is the total number of samples, K is 5000 for example. During the collection, the corn seeds are dropped one by one in a random posture by a vibrating seed metering plate, covering postures such as vertical long axis, horizontal long axis, tilt, and different orientations of the embryo surface. The dataset is divided into training set, validation set and test set in a ratio of 8:1:1, containing 4000, 500 and 500 paired samples respectively. The partitioning process is carried out on a sample-by-sample basis. All time data of the same seed belong to only one subset to avoid data with adjacent time appearing in both the training set and the test set at the same time, which would cause data leakage.

[0041] S6. Based on the signal attitude pairing supervised dataset, iterative attitude mapping model is used to establish a nonlinear mapping relationship between occlusion signal features and three-dimensional attitude parameters of corn seeds, and the trained attitude mapping model is obtained.

[0042] In one specific embodiment, performing S6 includes the following steps: A pose mapping model is constructed using a deep learning model. The input of the pose mapping model is the occlusion signal feature matrix, and the output is three-dimensional pose parameters. The occlusion signal feature matrix in the signal pose pairing supervision dataset is used as input data and the corresponding 3D pose label is used as output data to iteratively train the pose mapping model. In each training iteration, the error between the 3D pose parameters output by the pose mapping model and the 3D pose label is calculated, and the parameters of the pose mapping model are updated according to the error through the backpropagation algorithm to obtain the trained pose mapping model.

[0043] Specifically, the construction of the attitude mapping model is based on the structure of its input and output data. The input of the model is the occlusion signal feature matrix constructed in S4, with dimensions T×M×N, where T is the number of sampling times, M is the number of transmission channels, and N is the number of reception channels. The output of the model is three-dimensional attitude parameters, which are consistent with the data format of the attitude labels in S5, namely, quaternions representing spatial rotational attitude. The output dimension is 7-dimensional, consisting of a coordinate vector d representing the centroid position of the seed. Since the occlusion signal feature matrix contains both spatial and temporal information, the pose mapping model employs a deep learning structure of cascaded convolutional networks and gated recurrent networks, divided into a spatial feature extraction part, a temporal feature modeling part, and a pose regression output part. The spatial feature extraction part consists of three cascaded convolutional units. Each convolutional unit sequentially contains a 2D convolutional layer, a batch normalization layer, and a modified linear unit activation layer. The kernel size of each of the three convolutional layers is 3×3, with output channels of 16, 32, and 64 respectively. A 2×2 max-pooling layer is followed after the first and third convolutional layers. The spatial feature extraction part independently processes the M×N matrix slices at each sampling time, mapping the 8×8 matrix at each time step into a 64-dimensional matrix. The spatial feature vector is obtained by taking a matrix slice at T time points, which corresponds to a spatial feature sequence of length T and dimension 64. For example, when the matrix slice is 8×8, the first convolutional layer outputs an 8×8×16 feature map, which becomes 4×4×16 after pooling; the second convolutional layer outputs 4×4×32; and the third convolutional layer outputs 4×4×64, which becomes 2×2×64 after pooling. After flattening, a 256-dimensional vector is obtained, which is then compressed into a 64-dimensional spatial feature vector by a fully connected layer with an output dimension of 64. The temporal feature modeling part adopts a two-layer gated recurrent unit network, with the hidden state dimension of each layer set to 128. The gated recurrent unit selectively retains and forgets the occlusion features of historical time points through update gates and reset gates, enabling the model to utilize all occlusion evolution information from the seed entering the detection region to the current time point. After receiving the spatial feature sequence, the gated recurrent unit network outputs a 128-dimensional temporal feature vector corresponding to each time point. The attitude regression output consists of two fully connected layers. The first fully connected layer has an output dimension of 64 and uses modified linear unit activation. The second fully connected layer has an output dimension of 7, corresponding to 4-dimensional quaternions and 3-dimensional centroid coordinates. After output, the quaternion part is normalized to satisfy the unit quaternion constraint. The attitude parameters output by the model are denoted as follows: .

[0044] Model training was performed using a signal-pose pairing supervised dataset built with S5. The data foundation consists of a training set containing 4000 paired samples, and a validation set and a test set each containing 500 paired samples. During training, the occlusion signal feature matrix of the training set samples is used. The input data is fed into the attitude mapping model, and the predicted attitude parameters are obtained through forward propagation. and with pose labels in the same paired samples The error is calculated as the monitoring target. The loss function is a weighted sum of the rotation attitude error term and the position error term. The rotation attitude error term is in the form of a quaternion inner product, and the rotation attitude error at the k-th sampling time is... ,in This represents the inner product of the predicted quaternion and the label quaternion. This form is invariant to the duality of the quaternion's sign indicating the same rotation. The position error term uses the mean squared error form, which is the mean of the squares of the differences between the predicted centroid coordinates and the label centroid coordinates. The loss function for a single sample is... ,in The weighting coefficient for the position error term is used to balance the dimensions and numerical magnitudes of the rotational attitude error and the position error, for example... The centroid coordinates are set to 0.1, and normalized to a horizontal dimension of 30 mm for the detection region before being substituted into the loss function, ensuring they are within a comparable range to the dimensionless rotational attitude error. In each iteration, a batch of 32 paired samples is randomly selected from the training set. The batch loss is the average of the losses of all samples within that batch. The gradient of the batch loss with respect to the parameters of each model layer is then calculated using backpropagation. An adaptive moment estimation optimizer is used to update the model parameters. The initial learning rate of the adaptive moment estimation optimizer is set to 0.001, the exponential decay rate of the first-order moment estimation is set to 0.9, and the exponential decay rate of the second-order moment estimation is set to 0.999. The parameter update expression is as follows: ,in For learning rate, This is the bias correction value for the first moment estimate of the gradient. This is the bias correction value for the second moment estimate of the gradient. To prevent division by zero, the constant is set to a value of The learning rate employs a segmented decay strategy, decreasing to half its original value after every 50 training epochs. The total number of training epochs is set to 200. Simultaneously, after each epoch, the validation loss is calculated using the validation set. An early stopping strategy is adopted: training stops when the validation loss no longer decreases for 20 consecutive epochs, and the model parameters corresponding to the lowest validation loss are saved as the parameters of the trained pose mapping model. The trained pose mapping model undergoes performance verification on a test set. For example, the verification pass conditions are a pose angle deviation of no more than 2 degrees and a centroid position error of no more than 0.5 mm corresponding to the rotation pose error. Once the conditions are met, the model is deployed to the embedded computing platform of the seeding control terminal for online inference calls by the S7.

[0045] S7. During sowing operations, collect real-time occlusion signals, construct a real-time occlusion signal feature matrix, input it into the trained posture mapping model, and output the three-dimensional posture parameters of the corn seeds for sowing control and sensor calibration.

[0046] In one specific embodiment, performing S7 includes the following steps: During the sowing process, when corn seeds pass through the detection area, the composite light intensity response signal after being blocked by the corn seeds is collected in real time by the photoelectric receiving array; The composite light intensity response signal is sequentially amplified, filtered, and frequency separated to extract the occlusion response signal corresponding to each transmission channel, and then organized into a real-time occlusion signal feature matrix according to the time series. The real-time occlusion signal feature matrix is ​​input into the trained pose mapping model, and the three-dimensional pose parameters corresponding to the corn seed are output after the model performs forward calculation. The three-dimensional attitude parameters are sent to the seeding control execution unit, and the seeding direction of the seed metering device or the angle of the seed guide tube is adjusted according to the three-dimensional attitude parameters to achieve directional seeding control; The three-dimensional attitude parameters are sent to the sensor calibration unit and used as reference true values ​​for signal calibration of the white light emitting array and the photoelectric receiving array.

[0047] Specifically, during the sowing process, corn seeds are discharged by a seed metering device and pass through the detection area along the seed guide tube. A photoelectric receiving array collects the composite light intensity response signal after the seeds block the light. This acquisition process is completely identical to the S3 process in terms of hardware structure and signal form; that is, each of the N photoelectric receiving units outputs a photocurrent signal containing M sideband frequency components of the transmission channels, forming N composite light intensity response signals. The difference is that this stage no longer connects to the three-dimensional attitude calibration system; signal acquisition is completed independently by the multi-frequency modulated white light array detection device. When a seed enters and leaves the detection area, it blocks the entrance and exit trigger light paths, respectively. The entry trigger signal starts the recording of the current acquisition data, and the exit trigger signal marks the end of the current acquisition. The data between the two trigger signals is the signal segment corresponding to the entire process of the seed passing through the detection area. The duration of the signal segment corresponding to a single seed is 10 to 20 milliseconds, corresponding to 1000 to 2000 sampling times at a sampling frequency of 100 kHz.

[0048] The construction of the real-time occlusion signal feature matrix involves sequentially executing a processing chain for the composite light intensity response signal that is completely identical to the parameters in the training data acquisition stage. Specifically, each composite light intensity response signal is amplified by a preamplifier module with a transimpedance amplification structure and then sent to a frequency-selective filtering module. The N×M center frequencies are the upper sideband frequencies of each channel. The bandpass filter separates the M×N obstruction response signals, which are then coherently processed with the reference carrier signal at the corresponding sideband frequency and filtered by a 4th-order Butterworth low-pass filter with a cutoff frequency of 100 kHz to obtain the M×N baseband signals. Subsequently, the root mean square effective value of each baseband signal is extracted by a sliding window with a window length of 0.5 ms and corresponding to 50 sampling points. The normalized value is then normalized with the unobstructed reference effective value of the channel combination as the denominator to obtain the normalized obstruction coefficient with a value range of 0 to 1. The M×N normalized obstruction coefficients at each sampling time are filled into an M×N matrix according to the transmit channel index and the receive channel index and stacked along the time dimension to generate a real-time obstruction signal feature matrix with dimension T×M×N. Maintaining consistent parameters across all stages of the processing chain with those of the training stage is a prerequisite for consistent input distribution. Inconsistent parameters at any stage will cause a shift in the distribution of real-time features from that of the training features, thereby compromising the applicability of the mapping relationship learned by S6. Therefore, the unobstructed baseline effective value, sliding window length, and filter parameters used for normalization are directly inherited from the configuration of the training data acquisition stage.

[0049] The real-time occlusion signal feature matrix is ​​input into the trained pose mapping model. The model loads the parameters saved in S6 and outputs the three-dimensional pose parameters corresponding to the corn seed after forward computation. The forward computation process is a one-way execution during the training phase and does not involve parameter updates. That is, the spatial feature extraction part extracts a 64-dimensional spatial feature vector from the M×N matrix slice at each time step. The two-layer gated recurrent unit network of the temporal feature modeling part recursively obtains a 128-dimensional temporal feature vector at each time step along the time dimension. The pose regression output part outputs 7-dimensional pose parameters, of which the 4-dimensional quaternion represents the spatial rotation pose of the seed after modulus normalization. The 3-dimensional coordinate vector is multiplied by the horizontal dimension of the detection area of ​​30 mm to restore the centroid position in millimeters. For example, the attitude mapping model on the embedded computing platform takes less than 1 millisecond for a single forward computation. After the seed leaves the detection area and the trigger signal arrives, the forward computation of the complete matrix is ​​completed within 1 millisecond, and the attitude sequence of the entire process of the seed is output. Alternatively, a time-by-time incremental inference method can be adopted, that is, the gated recurrent unit network outputs the attitude parameters at the current moment as soon as it receives a new matrix slice, so that the attitude output is synchronized with the seed falling. The output quaternions can be further converted into Euler angles in the form of pitch angle, yaw angle, and roll angle, and combined with the centroid position to determine the seed embryo orientation and major axis direction.

[0050] The three-dimensional attitude parameters are sent to the sowing control execution unit. The sowing control execution unit adjusts the sowing direction of the seed metering device or the angle of the seed guide tube according to the three-dimensional attitude parameters to achieve directional sowing control. Specifically, the sowing control execution unit compares the output attitude parameters with the preset target soil entry attitude. The target soil entry attitude is set, for example, with the embryo surface facing down and the long axis aligned with the sowing row direction. The attitude deviation angle between the current attitude and the target attitude is calculated. When the attitude deviation angle is greater than the preset adjustment threshold, the sowing control execution unit drives the attitude adjustment mechanism at the outlet of the seed metering device. For example, the attitude adjustment mechanism is a servo-driven deflection section at the end of the seed guide tube. The servo adjusts the angle of the deflection section according to the direction and magnitude of the attitude deviation angle, so that the attitude of the seed when leaving the seed guide tube is corrected to the target soil entry attitude. The adjustment threshold is set, for example, to 10 degrees. When the attitude deviation angle is less than 10 degrees, the attitude is considered qualified and no adjustment action is triggered to avoid mechanical vibration caused by frequent action of the execution mechanism.

[0051] The three-dimensional attitude parameters are sent to the sensor calibration unit, which uses them as a reference truth value to calibrate the white light emitting array and the photoelectric receiving array. Specifically, in calibration mode, the sensor calibration unit controls a standard attitude seed sample to pass through the detection area, compares the three-dimensional attitude parameters output by the attitude mapping model with the set attitude of the standard attitude seed sample, and records the normalized occlusion coefficient of each channel combination under the set attitude. When the deviation between the measured normalized occlusion coefficient of a certain channel combination and the historical value under the same set attitude exceeds a preset threshold, it is determined that the link corresponding to that channel combination has experienced gain drift. The preset threshold is exemplarily set to 0.05. Subsequently, the effective value of the unobstructed reference is recalculated based on the current unobstructed output of the channel combination, and the normalized denominator is updated. This ensures that the occlusion response reference of each channel is periodically corrected with emission attenuation and receiving gain drift, guaranteeing that the real-time occlusion signal feature matrix and the training data feature distribution remain consistent during long-term operation.

[0052] Please see Figure 3 The following describes a corn seed attitude estimation system according to an embodiment of this application. The corn seed attitude estimation system includes: The acquisition and arrangement module is used to arrange a white light emitting array and a photoelectric receiving array opposite each other on both sides of the corn seed falling path to form a detection area. Each white light emitting unit of the white light emitting array constitutes an independent emission channel, and each photoelectric receiving unit of the photoelectric receiving array constitutes an independent receiving channel. The multi-frequency modulation module is used to generate modulation signals with different frequencies through the modulation signal generation module and load them onto each white light emitting unit, outputting a multi-frequency modulated white light signal with a unique frequency domain identifier. The signal separation module is used to block the multi-frequency modulated white light signals in different detection optical paths when corn seeds pass through the detection area. The photoelectric receiving array outputs a composite light intensity response signal, which is amplified, filtered and frequency-selected to separate the blocking response signals of each transmission channel. The signal integration module is used to structure and organize the occlusion response signals of each transmission channel and each receiving channel in the time series, and generate an occlusion signal feature matrix that can characterize the dynamic occlusion changes of the seed. The data establishment module is used to perform temporal and spatial synchronization registration between the occlusion signal feature matrix and the attitude labels generated by the 3D attitude calibration system to construct a signal attitude pairing supervision dataset. The model training module is used to iterate the pose mapping model based on the signal pose pairing supervised dataset, establish the nonlinear mapping relationship between occlusion signal features and the three-dimensional pose parameters of corn seeds, and obtain the trained pose mapping model. The sowing control module is used to collect real-time occlusion signals during sowing operations, construct a real-time occlusion signal feature matrix and input it into the trained posture mapping model, and output the three-dimensional posture parameters of the corn seeds for sowing control and sensor calibration.

[0053] Through the synergistic cooperation of the above components, the direct inference and output of the three-dimensional attitude parameters of corn seeds from the composite light intensity response signal of the multi-frequency modulated white light array is achieved, wherein: The acquisition and deployment module establishes the hardware detection optical path foundation. Through a dual-sided through-beam white light transmitter and photoelectric receiver array, independent transmit and receive channels are divided, forming a three-dimensional light field detection area covering the trajectory of falling corn seeds. This provides a hardware platform for capturing differentiated optical path occlusion under different seed postures, overcoming the limitation of single-optical-path sensing which can only acquire one-dimensional occlusion information. The multi-frequency modulation module assigns a unique frequency domain identifier to each transmit channel, modulating the optical signal. Spectrum shifting is achieved through carrier modulation. The receiver can distinguish the signals from each channel through frequency domain filtering, overcoming the time resolution sacrifice of time-division multiplexing schemes and realizing multi-channel optical paths. Parallel synchronous acquisition fully preserves the continuous occlusion timing information of the entire high-speed seed descent process; the signal separation module completes photoelectric conversion, signal amplification, and frequency domain filtering and demodulation, separating the independent occlusion response electrical signal of each receiving and receiving path from the multi-channel mixed composite light intensity signal, removing ambient light and circuit noise interference, and extracting the effective light intensity change caused only by the occlusion of the seed entity; the signal integration module normalizes and smooths the effective value of each occlusion signal, constructs an occlusion signal feature matrix in a three-dimensional structure of time, transmission channel, and receiving channel, and uniformly encapsulates the spatial occlusion distribution and temporal dynamic changes. The system features physical occlusion phenomena transformed into standardized tensor data readable by the model. The data establishment module builds a complete calibration dataset link, obtains true pose values ​​through external 3D pose calibration equipment, and completes synchronous registration of signal features and pose labels in both time and space. It matches the occlusion matrix with the corresponding 3D pose one by one, solving the accuracy bottleneck of traditional photoelectric sensors that lack systematic calibration and rely on manual geometric model mapping. The model training module builds a pose mapping model based on a convolutional + GRU temporal deep learning network. It iteratively optimizes network parameters with paired supervised datasets, automatically learning the nonlinear correlation between multi-dimensional occlusion features and the rotation, position, and 3D pose of corn seeds. It achieves end-to-end pose regression without the need for manual threshold setting or geometric fitting. The sowing control module undertakes online inference and field application functions. During sowing operations, it reproduces the complete signal processing flow in real time to generate a real-time feature matrix. After training, the model quickly outputs the 3D pose parameters of the seeds. On the one hand, it feeds back to the seeding and guiding actuators to complete directional sowing control. On the other hand, it serves as a true value benchmark to periodically correct the optical path gain of the photoelectric array, offsetting detection errors caused by light source attenuation and sensor drift, and ensuring the long-term operational stability of the equipment.

[0054] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the corn seed attitude estimation method described above.

[0055] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the methods and systems described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0056] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for estimating the attitude of maize seeds, characterized in that, include: S1. A detection area is formed by arranging a white light emitting array and a photoelectric receiving array opposite each other on both sides of the corn seed falling path. Each white light emitting unit of the white light emitting array constitutes an independent emission channel, and each photoelectric receiving unit of the photoelectric receiving array constitutes an independent receiving channel. S2. The modulation signal generation module generates modulation signals with different frequencies and loads them onto each white light emitting unit, outputting a multi-frequency modulated white light signal with a unique frequency domain identifier. S3. When corn seeds pass through the detection area, they block the multi-frequency modulated white light signals in different detection optical paths. The photoelectric receiving array outputs a composite light intensity response signal, which is then amplified, filtered, and frequency-selected to separate the blocking response signals of each transmission channel. S4. The occlusion response signals of each transmit channel and each receive channel in the time series are structured and organized to generate an occlusion signal feature matrix that can characterize the dynamic occlusion changes of the seed. This includes: demodulating the occlusion response signals, extracting the baseband signal corresponding to the occlusion changes of each channel by coherent operation with the reference carrier signal and low-pass filtering; extracting the effective value of the baseband signal to obtain the occlusion response intensity value between each transmit channel and each receive channel; and organizing the occlusion response intensity values ​​of all transmit channels and all receive channels at each sampling time in time sequence to construct the occlusion signal feature matrix. S5. Perform time and space synchronization registration between the occlusion signal feature matrix and the attitude labels generated by the three-dimensional attitude calibration system to construct a signal attitude pairing supervision dataset. S6. Based on the signal attitude pairing supervision dataset, iterate the attitude mapping model to establish a nonlinear mapping relationship between the occlusion signal features and the three-dimensional attitude parameters of the corn seed, and obtain the trained attitude mapping model. S7. During sowing operations, collect real-time occlusion signals, construct a real-time occlusion signal feature matrix, input it into the trained posture mapping model, and output the three-dimensional posture parameters of the corn seeds for sowing control and sensor calibration.

2. The method according to claim 1, characterized in that, S1 includes: A white light emitting array is arranged on the first side of the corn seed falling path, the white light emitting array comprising multiple white light emitting units arranged in a horizontal array along the seed falling direction; A photoelectric receiving array is arranged on the second side of the corn seed falling path, and the photoelectric receiving array includes multiple photoelectric receiving units corresponding to the white light emitting unit; The white light emitting array and the photoelectric receiving array are arranged opposite each other, so that each white light emitting unit and each photoelectric receiving unit form multiple pairs of oppositely arranged emitting and receiving correspondences in space, and a detection area for corn seeds to fall freely is formed between them.

3. The method according to claim 1, characterized in that, S2 includes: The modulation signal generation module generates multiple periodic modulation signals, and the modulation frequencies of each transmission channel are different and meet the preset frequency interval constraints. The signal shaping and filtering module performs waveform shaping, amplitude adjustment and DC bias superposition on the modulation signal to keep the white light emitting unit in a stable light-emitting state. The carrier modulation module loads the processed modulation signal onto a high-frequency carrier signal to form a composite modulation signal, thereby achieving spectrum shifting. The white light driving module controls the driving current of the white light emitting unit according to the composite modulation signal, and outputs a multi-frequency modulated white light signal.

4. The method according to claim 1, characterized in that, S3 includes: When corn seeds enter the detection area between the white light emitting array and the photoelectric receiving array, they simultaneously block multiple spatial detection optical paths between each emitting unit of the white light emitting array and each receiving unit of the photoelectric receiving array. Each photoelectric receiving unit receives the multi-frequency modulated white light signal that arrives after being blocked by corn seeds, converts the optical signal into an electrical signal, and outputs a composite light intensity response signal containing frequency domain components of multiple transmission channels; The composite light intensity response signal is amplified by the preamplifier module and then sent to the frequency selective filter module. The modulation frequency components corresponding to each transmission channel are extracted by multiple bandpass filters with different center frequencies, and the blocking response signal of each transmission channel is obtained.

5. The method according to claim 1, characterized in that, S5 includes: The three-dimensional attitude label of the corn seed at each moment during the process of passing through the detection area is obtained by a three-dimensional attitude calibration system. The three-dimensional attitude calibration system is an external calibration device independent of the multi-frequency modulation white light array detection device. The attitude label includes the spatial rotation attitude parameters of the corn seed. Using the triggering events of corn seeds entering and leaving the detection area as the time reference, the sampling times of each sampling moment in the occlusion signal feature matrix and the attitude labels output by the three-dimensional attitude calibration system are synchronously aligned on the time axis, so that the occlusion signal features of each sampling moment and the attitude labels of the same moment form a one-to-one correspondence. The coordinate system of the detection area and the coordinate system of the three-dimensional attitude calibration system are spatially unified and calibrated. The attitude label is transformed to the spatial reference coordinate system where the detection area is located through coordinate transformation, so that the occlusion signal features and the attitude label are in the same coordinate system. The occlusion signal feature matrix after time and space synchronization registration is paired with the corresponding attitude label. Each corn seed passes through the detection area and corresponds to an occlusion signal feature matrix sample and an attitude label sample. The two constitute a paired sample. All paired samples are aggregated to construct a signal attitude pairing supervised dataset.

6. The method according to claim 1, characterized in that, S6 includes: A pose mapping model is constructed, which adopts a deep learning model. The input of the pose mapping model is the occlusion signal feature matrix, and the output is three-dimensional pose parameters. The occlusion signal feature matrix in the signal pose pairing supervision dataset is used as input data, and the corresponding three-dimensional pose label is used as output data to iteratively train the pose mapping model. In each iteration of training, the error between the three-dimensional pose parameters output by the pose mapping model and the three-dimensional pose label is calculated, and the parameters of the pose mapping model are updated according to the error through the backpropagation algorithm to obtain the trained pose mapping model.

7. The method according to claim 1, characterized in that, S7 includes: During the sowing process, when corn seeds pass through the detection area, the composite light intensity response signal after being blocked by the corn seeds is collected in real time by the photoelectric receiving array; The composite light intensity response signal is sequentially amplified, filtered, and frequency separated to extract the occlusion response signal corresponding to each transmission channel, and then organized into a real-time occlusion signal feature matrix according to the time sequence. The real-time occlusion signal feature matrix is ​​input into the trained pose mapping model, and the three-dimensional pose parameters corresponding to the corn seed are output after the model performs forward calculation. The three-dimensional attitude parameters are sent to the seeding control execution unit, and the seeding direction of the seed metering device or the angle of the seed guide tube is adjusted according to the three-dimensional attitude parameters to achieve directional seeding control; The three-dimensional attitude parameters are sent to the sensor calibration unit, and the three-dimensional attitude parameters are used as reference true values ​​to perform signal calibration on the white light emitting array and the photoelectric receiving array.

8. A maize seed attitude estimation system, used to implement the maize seed attitude estimation method as described in any one of claims 1 to 7, characterized in that, include: The acquisition and arrangement module is used to arrange a white light emitting array and a photoelectric receiving array opposite each other on both sides of the corn seed falling path to form a detection area. Each white light emitting unit of the white light emitting array constitutes an independent emission channel, and each photoelectric receiving unit of the photoelectric receiving array constitutes an independent receiving channel. The multi-frequency modulation module is used to generate modulation signals with different frequencies through the modulation signal generation module and load them onto each white light emitting unit, outputting a multi-frequency modulated white light signal with a unique frequency domain identifier. The signal separation module is used to block the multi-frequency modulated white light signals in different detection optical paths when corn seeds pass through the detection area. The photoelectric receiving array outputs a composite light intensity response signal, which is amplified, filtered, and frequency-selected to separate the blocking response signals of each transmission channel. The signal integration module is used to structure and organize the occlusion response signals of each transmission channel and each receiving channel in the time series, and generate an occlusion signal feature matrix that can characterize the dynamic occlusion changes of the seed. The data establishment module is used to perform time and space synchronization registration between the occlusion signal feature matrix and the attitude labels generated by the three-dimensional attitude calibration system to construct a signal attitude pairing supervision dataset. The model training module is used to iterate the pose mapping model based on the signal pose pairing supervision dataset, establish a nonlinear mapping relationship between occlusion signal features and three-dimensional pose parameters of corn seeds, and obtain the trained pose mapping model. The sowing control module is used to collect real-time occlusion signals during sowing operations, construct a real-time occlusion signal feature matrix and input it into the trained posture mapping model, and output the three-dimensional posture parameters of the corn seeds for sowing control and sensor calibration.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements a corn seed attitude estimation method as described in any one of claims 1 to 7.

Images