Method for reservoir computing and motion direction recognition based on ferroelectric polarization control response

Abstract

The present invention discloses a storage pool calculation and motion direction recognition method based on ferroelectrically controlled polarization response, comprising the following steps: S1, encoding the temporal dynamic image data of the moving target to be identified into a multi-bit time-series optical pulse signal, wherein the temporal dynamic image data of the moving target to be identified consists of multiple consecutive frames; S2, inputting the generated time-series optical pulse signal into a ferroelectrically controlled polarization-sensitive photosynaptic device for stimulation; under the driving of time-series optical stimulation, the photocurrent output by the polarization-sensitive photosynaptic device exhibits nonlinear dynamic evolution and reaches a steady-state response with the input of the pulse sequence; the steady-state photocurrent value or equivalent conductance value after the evolution is completed is used as the state output of the physical storage pool; S3, the state output of the physical storage pool is fed into a fully connected network to realize the motion target pattern classification and recognition. This invention provides a low-power, high-real-time, and highly interference-resistant edge artificial vision method.

Description

Technical Field

[0001] This invention belongs to the field of photoelectric detection technology, specifically relating to a method for calculating the direction of motion of a storage cell based on ferroelectric modulation polarization response. Background Technology

[0002] Motion direction recognition refers to detecting motion in an object or scene and accurately determining the direction of that motion. Essentially, it involves perceiving and discriminating the directional attributes of motion vectors. This requires not only sensing the existence of motion but also effectively encoding and distinguishing the temporal evolution characteristics of the motion. Currently, the technical solutions for motion direction recognition mainly fall into the following two categories: One approach is based on traditional frame-based visual sensors and software algorithms. These methods typically rely on high-frame-rate cameras to acquire continuous images and use optical flow, background modeling, feature point tracking, or deep learning models such as a combination of convolutional and recurrent neural networks to calculate the displacement vectors of pixels or regions between adjacent frames. While these approaches can achieve high recognition accuracy in clear and stable visual environments, their processing often involves the continuous acquisition, storage, and transmission of large amounts of image data and relies on complex numerical calculations, resulting in high system latency and power consumption, making it difficult to achieve real-time, efficient processing on resource-constrained edge devices. Furthermore, their performance often degrades significantly in scenarios with low visibility, drastic lighting changes, or complex background interference.

[0003] The second approach utilizes emerging neuromorphic vision sensors and brain-like computing architectures. Inspired by biological vision systems, these solutions directly output asynchronous pulse event streams related to changes in light intensity using dynamic vision sensors or pulse cameras, thus introducing temporal information at the sensor level. This type of data features high temporal resolution, low redundancy, and low power consumption. Combined with models such as spiking neural networks or reservoir computing, it can achieve efficient motion direction perception to a certain extent. However, most existing neuromorphic vision solutions still rely primarily on backend networks to extract and classify features from event streams. Their sensors themselves typically lack the physical nonlinear mapping and memory capabilities for complex temporal patterns, making it difficult to effectively separate and encode motion temporal features at the device level. Furthermore, the perception and processing capabilities of existing solutions remain limited for complex motion patterns involving polarization correlation or multidimensional optical information.

[0004] Therefore, how to overcome the limitations of traditional motion direction recognition systems in terms of dynamic feature encoding and memory capabilities at the sensor end in low visibility environments such as rain and snow, as well as other complex visual environments such as strong light reflection, dynamic occlusion, and background noise interference, and provide a novel integrated architecture of perception, storage, and computing that can directly fuse multi-dimensional light information at the sensor end and utilize the intrinsic nonlinear dynamics and memory characteristics of the device under polarization response to achieve temporal feature mapping and storage, thereby realizing a motion direction recognition method with high robustness and real-time response, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method for calculating the direction of motion of a storage pool based on ferroelectric polarization response, addressing the problems in the prior art.

[0006] Therefore, the above-mentioned objectives of the present invention are achieved through the following technical solutions: A method for calculating and identifying the motion direction of a storage cell based on ferroelectric polarization response includes the following steps: S1, the time-series dynamic image data of the moving target to be identified is encoded into a multi-bit time-series light pulse signal, and the time-series dynamic image data of the moving target to be identified is a continuous multi-frame image; S2, The time-series optical pulse signal generated in step S1 is input into the ferroelectrically modulated polarization-sensitive photoelectric synapse device for stimulation; Under the time-series optical stimulation, the photocurrent output by the polarization-sensitive photoelectric synapse device exhibits nonlinear dynamic evolution and reaches a steady-state response with the input of the pulse sequence; The steady-state photocurrent or equivalent conductance value after the evolution is completed will be used as the state output of the physical reservoir. S3. The state output of the physical reservoir described in step S2 is input into the subsequent fully connected neural network to complete the classification and recognition of the motion pattern of the moving target.

[0007] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions: As a preferred technical solution of the present invention: In step S1, the moving target is a vehicle, the time-series dynamic image of the moving target is a 36×36 pixel grayscale image from a top-down perspective, the road layout consists of a pair of mutually orthogonal corridor-type roads with a width of 12 pixels, the vehicle is represented by a 3×3 pixel high grayscale block, each data sample consists of four consecutive frames of images, the four consecutive 36×36 pixel vehicle motion images are converted into 4-bit, 1296 time-series optical pulse signals, in the 4-bit time-series optical pulse signal "1" represents an excitation optical pulse in a specific polarization direction, the illumination pulse time is 1s, "0" represents no illumination, forming a pulse sequence, which is input to the ferroelectric control polarization sensitive photoelectric synapse device.

[0008] As a preferred technical solution of the present invention: In step S1, the moving target temporal dynamic image data includes eight different motion directions and is composed of temporal data samples consisting of multiple consecutive grayscale images. The eight different directions of movement include: going straight, turning left, turning right, and making a U-turn when entering from the right side of the road, and going straight, turning left, turning right, and making a U-turn when entering from the left side of the road.

[0009] As a preferred technical solution of the present invention: In step S2, while performing light stimulation, an initialization voltage pulse with an absolute value greater than the coercive field of the ferroelectric thin film is applied to the gate of the ferroelectric modulated polarization sensitive photoelectric synapse device to regulate the channel and interface state; after each timing signal processing, a reverse gate voltage is applied to erase the ferroelectric polarization state to ensure that the initial state of the device is the same before the next timing signal processing.

[0010] As a preferred embodiment of the present invention: in step S2, the ferroelectrically modulated polarization-sensitive photosynaptic device includes, The polarization light sensitive unit includes at least one photosynaptic ferroelectric transistor, wherein the two-dimensional semiconductor channel layer of the photosynaptic ferroelectric transistor is composed of a two-dimensional semiconductor material with optical anisotropy, and its ferroelectric functional layer is composed of a ferroelectric material. A polarization light modulation module is disposed on the light receiving path of the photosynaptic ferroelectric transistor to provide the photosynaptic ferroelectric transistor with an adjustable polarization direction; The ferroelectric functional layer serves as the presynapse, and the two-dimensional semiconductor channel layer serves as the postsynapse. The weights of the channel obtained under ferroelectric modulation are updated and adjusted under electrical and optical stimulation to process polarization visual information. Under the modulation of ferroelectric polarization, the device exhibits enhanced nonlinear response characteristics to illumination conditions corresponding to the direction of strongest polarization response, thereby achieving deep fusion of polarization information and nonlinear synaptic dynamics.

[0011] As a preferred embodiment of the present invention: the photosynthetic ferroelectric transistor includes: The substrate is a silicon substrate covered with a silicon dioxide layer; The gold electrode is located on the silicon dioxide layer; The ferroelectric functional layer, covering the gold electrode and part of the substrate, is a polyvinylidene fluoride ferroelectric polymer film; A two-dimensional semiconductor channel layer, placed on the ferroelectric functional layer, is an anisotropic palladium diselenide channel layer with a thickness of 40 nanometers. The chromium-gold composite electrode forms ohmic contacts with both ends of the two-dimensional semiconductor channel layer. The lower layer of the chromium-gold composite electrode is a chromium layer with a thickness of 10 nanometers, and the upper layer is a gold layer with a thickness of 50 nanometers. The polarization light modulation module irradiates the photosynaptic ferroelectric transistor, providing polarized light stimulation to the two-dimensional semiconductor channel layer. By applying a gate voltage higher than the coercive field of the ferroelectric material on the chromium-gold composite electrode, the polyvinylidene fluoride ferroelectric polymer film is fully polarized, thereby injecting electrons or holes into the palladium diselenide channel layer, achieving regulation of carrier concentration and photosynaptic characteristics.

[0012] As a preferred technical solution of the present invention: In step S2, by adjusting the polarization direction of the incident light, the state space distribution of the physical reservoir is actively controlled. Under the 90° polarization direction, the ferroelectric control polarization-sensitive photoelectric synapse device enables the final state current corresponding to different input sequences to form a high separation distribution in the state space, thereby achieving its high classification performance in the time sequence recognition task.

[0013] As a preferred technical solution of the present invention: In step S3, the time-series optical pulse signal under the motion direction encoded in step S1 is matched and labeled with the physical reservoir state output obtained in step S2 to construct a supervised learning dataset for training the readout layer neural network, specifically including: S301, Encode each frame of moving target time-series dynamic image data with known motion direction into a multi-bit time-series light pulse signal to obtain multiple sets of input sequences; S302, each set of input sequences is input to the ferroelectric photoelectric synapse device, repeated measurements are performed, and one of the multiple photocurrent response final state values ​​is selected as the physical reservoir state output value corresponding to the input sequence; S303, construct a set of state vectors from the physical reservoir state output values ​​corresponding to the multiple sets of input sequences, and associate them with the corresponding motion direction labels, thereby forming a supervised learning dataset for training the readout layer neural network.

[0014] As a preferred embodiment of the present invention: the readout layer neural network is a fully connected neural network, comprising: The input layer, whose number of nodes is the same as the dimension of the state vector, is used to receive the state vector composed of the state output values ​​of the physical reserve pool. At least one hidden layer is used for nonlinear transformation and feature extraction of the state vector; The output layer, with the same number of nodes as the number of motion direction categories to be classified, is used to output the classification results.

[0015] As a preferred technical solution of the present invention: using a trained fully connected neural network, the time-series dynamic image data of the moving target in the motion direction to be identified in step S1 is encoded into a 4-bit time-series light pulse signal, generating a total of 1296 different input sequences. In step S3, the input layer contains 1296 neurons, which correspond one-to-one with the state output dimension of the physical reservoir, and fully receives the 1296-dimensional state feature vector output from step S2. The hidden layer contains 128 neurons, and the output layer contains 8 neurons, which correspond to 8 preset motion directions, output the probability value of each direction category, and take the category with the highest probability value as the final recognition result.

[0016] Compared with existing technologies, the temporal optical information processing and motion recognition method based on ferroelectric photoelectric synaptic physical reservoir of the present invention has the following beneficial effects: This invention combines ferroelectric materials with two-dimensional semiconductor materials (such as palladium diselenide, PdSe2) exhibiting optical anisotropy to construct a ferroelectrically modulated polarization-sensitive photoelectric synapse device. This device not only inherits the non-volatile modulation capability of ferroelectric materials over channel carriers but also incorporates the intrinsic polarization-selective absorption characteristics of two-dimensional materials. Therefore, under polarized light stimulation, the sensor directly generates a photoelectric response signal with time-dependent, nonlinear, and memory characteristics, providing a foundation for subsequent information processing.

[0017] This invention utilizes light pulses with different polarization directions and their occurrence times to encode timing information, and performs dynamic modulation through ferroelectric polarization, enabling the device to produce highly differentiated nonlinear responses to timing inputs with different polarizations. This polarization-timing joint encoding constructs the state space of the storage pool, improving its ability to extract and distinguish features from complex dynamic patterns such as vehicle motion. The polarization-sensitive photoelectric synapse device in this invention is both a sensor for sensing polarization-sequential optical signals and a computing node for generating a storage pool of features through nonlinear dynamic mapping. This reduces the data transmission and computation process in traditional vision systems, which involves step-by-step processing from image sensing, analog-to-digital conversion to digital processing units. It achieves low-latency, low-power motion direction recognition and improves the real-time response capability of the system.

[0018] This invention suppresses environmental scattering interference directly at the signal sensing stage through polarization resolution and nonlinear processing of the device, effectively extracting the intrinsic direction features of motion. Under optimal polarization 90° control, the recognition accuracy reaches 96.3%, which is significantly better than the traditional linear response mode, demonstrating its excellent performance in vehicle motion direction recognition under harsh visual conditions and complex environments.

[0019] The present invention relates to a temporal optical information processing and motion recognition method based on a ferroelectric photoelectric synaptic physical reservoir. Taking a ferroelectrically modulated polarization-sensitive photoelectric synaptic device as the core, it constructs a reservoir computing system based on polarization-temporal coding. For the first time, it realizes polarization-discriminated perception, nonlinear feature extraction and memory fusion of complex dynamic visual signals of vehicle motion direction at the device level. It provides a brand-new hardware solution for high real-time performance and strong anti-interference edge artificial vision system, which has great application prospects. Attached Figure Description

[0020] Figure 1 This is a diagram of the ferroelectric-modulated polarization-sensitive photoelectric synapse device and polarized light system of the present invention. In the attached figures, the substrate is 101; the chromium-gold composite electrode is 102; the ferroelectric functional layer is 103; the two-dimensional semiconductor channel layer is 104; the gold electrode is 105; the polarizer is 201; the half-wave plate is 202; and the light source is 300. Figure 2 The optical response process is shown in the diagrams for 0° and 90° polarized laser pulses. Figure 3 The diagram shows the corresponding photocurrent differentiation effect under 16 different input light pulse sequences at 90°. Figure 4 The diagram shows the effect of photocurrent differentiation under 16 different input light pulse sequences at 0°. Figure 5 A diagram illustrating the construction process of a physical memory based on polarization-sensitive ferroelectric devices (taking a straight line appearing on the left side of a vehicle as an example). Figure 6 This refers to a storage pool and a fully connected neural network structure. Figure 7 This is a graph showing the accuracy results for identifying the vehicle's direction of travel. Figure 8 This is a clustering diagram showing the differentiation of eight vehicle motion directions under 90° polarization. Figure 9 This is a clustering diagram showing the differentiation of eight vehicle motion directions under 0° polarization. Figure 10 The confusion matrix represents the classification of eight vehicle motion directions under 90° polarization. Figure 11 This is the confusion matrix for classifying the eight vehicle motion directions under 0° polarization. Detailed Implementation

[0021] The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

[0022] The present invention provides a method for calculating the direction of motion of a storage cell based on ferroelectric polarization response, comprising the following steps: S1: Convert the time-series dynamic image data of the moving target to be identified into a 4-bit time-series light pulse signal; S2: The 4-bit time-series optical pulse signal generated in step S1 is directly input into the ferroelectrically modulated polarization-sensitive photoelectric synapse device. Under the drive of time-series optical stimulation, the output photocurrent of the device exhibits a nonlinear dynamic response and its output photocurrent exhibits a nonlinear dynamic evolution. As the pulse sequence is input, it gradually reaches a final state response value. The final state photocurrent response shows significant distinguishability between different input sequences and is highly sensitive to the time sequence of the pulses. The final steady-state current value after the evolution is completed is extracted as the physical reservoir state output of the corresponding input sequence, thus completing the feature mapping from time-series input to high-dimensional state on the device. S3: Collect the state output of the physical reservoir, construct a dataset for training the classification network, and train a fully connected neural network using the dataset. The network is configured to identify and classify the corresponding vehicle movement direction based on the input physical reservoir state output.

[0023] In step S1, the moving target temporal dynamic image data includes temporal data samples consisting of eight different motion directions and multiple consecutive grayscale images. The eight types of vehicle movement directions include: straight, left turn, right turn, and U-turn when entering from the right side of the road, and straight, left turn, right turn, and U-turn when entering from the left side of the road.

[0024] The moving target time-series dynamic image is a 36×36 pixel grayscale image from a top-down perspective. The road layout consists of a pair of mutually orthogonal corridor-style roads with a width of 12 pixels. Vehicles are represented by 3×3 pixel high grayscale blocks. Each data sample consists of four consecutive frames.

[0025] In the 4-bit timed optical pulse signal, "1" represents an excitation optical pulse with a specific polarization direction, and "0" represents an excitation pulse with no light or orthogonal polarization direction, forming a pulse sequence that is input to a ferroelectrically modulated polarization-sensitive photoelectric synapse device.

[0026] As a preferred embodiment of the present invention: the ferroelectric-modulated polarization-sensitive photosynaptic device comprises, A polarization light sensitive unit includes at least one photosynaptic ferroelectric transistor, wherein the two-dimensional semiconductor channel layer of the photosynaptic ferroelectric transistor is composed of a two-dimensional semiconductor material with optical anisotropy, and the ferroelectric functional layer is composed of a ferroelectric material. A polarization light modulation module is disposed on the light receiving path of the photosynaptic ferroelectric transistor to provide the photosynaptic ferroelectric transistor with an adjustable polarization direction; The ferroelectric functional layer serves as the presynapse, and the two-dimensional semiconductor channel layer serves as the postsynapse. The weights of the channel obtained under ferroelectric modulation are updated and adjusted under electrical and optical stimulation to process polarization visual information. Under the modulation of ferroelectric polarization, the device exhibits enhanced nonlinear response characteristics to illumination conditions corresponding to the direction of strongest polarization response, thereby achieving deep fusion of polarization information and nonlinear synaptic dynamics.

[0027] Simultaneously with photostimulation, an initialization voltage pulse with an absolute value greater than the coercive field of the ferroelectric thin film is applied to the gate of the ferroelectrically modulated polarization-sensitive photosynaptic device to modulate the channel and interface states. After each timing signal processing, a reverse gate voltage is applied to erase the ferroelectric polarization state, ensuring that the device's initial state is the same before the next timing signal processing. The photosynaptic ferroelectric transistor includes: The substrate is a silicon substrate covered with a silicon dioxide layer; The gold electrode is located on the silicon dioxide layer; The ferroelectric functional layer, covering the gold electrode and part of the substrate, is a polyvinylidene fluoride ferroelectric polymer film; A two-dimensional semiconductor channel layer, placed on the ferroelectric functional layer, is an anisotropic palladium diselenide channel layer with a thickness of 40 nanometers. The chromium-gold composite electrode forms ohmic contacts with both ends of the two-dimensional semiconductor channel layer. The lower layer of the chromium-gold composite electrode is a chromium layer with a thickness of 10 nanometers, and the upper layer is a gold layer with a thickness of 50 nanometers. The polarization light modulation module is irradiated above the photosynaptic ferroelectric transistor to provide polarization light stimulation to the two-dimensional semiconductor channel layer; By applying a gate voltage higher than the coercive field of the ferroelectric material on the chromium-gold composite electrode, the polyvinylidene fluoride ferroelectric polymer film is fully polarized, thereby injecting electrons or holes into the palladium diselenide channel layer, achieving regulation of carrier concentration and photosynaptic properties.

[0028] In step S2, under the 90° polarization direction, the ferroelectric modulated polarization-sensitive photoelectric synapse device enables the final state current corresponding to different input sequences to form a high separation distribution in the state space. By adjusting the polarization direction of the incident light, the state space distribution of the physical reservoir is actively controlled, thereby achieving its high classification performance in the time sequence recognition task.

[0029] In S3, motion pattern classification and recognition based on the physical reservoir state output: The physical reservoir state obtained in step S2 is output and fed into a trained three-layer fully connected readout layer neural network to classify and identify the vehicle's motion direction. The structure of this readout layer neural network is as follows: Input layer: Contains 1296 neurons, which correspond one-to-one with the state output dimensions of the physical reservoir, and fully receive the 1296-dimensional state feature vector output by step S2.

[0030] Hidden layer: Contains 128 neurons and is fully connected to the input layer. This layer uses its trained weight parameters to perform nonlinear transformation and effective compression on the high-dimensional features of the input, extracting the key feature representations most relevant to motion direction discrimination.

[0031] Output layer: Contains 8 neurons, each corresponding to a preset vehicle movement direction category (straight, left turn, right turn, U-turn from the right side of the road, and straight, left turn, right turn, U-turn from the left side of the road). This layer outputs the probability of belonging to each movement direction category; finally, the direction corresponding to the highest probability value is determined as the recognition result.

[0032] The network is trained based on a specially constructed supervised learning dataset, and the specific construction steps are as follows: S301. Timing signal encoding: The time-series dynamic image data of the moving target with the motion direction marked in step S1 is encoded into a 4-bit time-series optical pulse signal through a preprocessing method based on reservoir calculation, generating a total of 1296 different input sequences.

[0033] S302. State output acquisition and randomization: For each set of input sequences, the ferroelectric photoelectric synapse device is repeatedly measured 10 times to obtain 10 final state values ​​of photocurrent response; one value is randomly selected from each set of measurement results as the physical reservoir state output value corresponding to that input sequence.

[0034] S303. Dataset Construction: Combine the state output values ​​corresponding to all input sequences into 1296 state vectors, and precisely match them with the corresponding vehicle motion direction labels to form the final supervised learning dataset. Similarly, for the input sequence consisting of multiple frames of images with unknown motion directions, the final state value of the photocurrent response of the time-series optical pulse signal is randomly selected from 10 repeated measurements as the physical reservoir state output value corresponding to the input sequence. Using this dataset, the readout layer neural network is trained using the error backpropagation algorithm. As a preferred technical solution of the present invention: In step S3, the number of input layer nodes of the fully connected neural network matches the dimension of the physical reservoir state output, and the number of output layer nodes is 8, corresponding to eight types of vehicle motion directions.

[0035] The fully connected neural network includes an input layer, at least one hidden layer, and an output layer. Its network structure is 1296×128×8, and it is built and implemented using the Python programming language in the TensorFlow and Keras deep learning framework.

[0036] This invention presents a method for calculating a storage cell and identifying its motion direction based on ferroelectrically modulated polarization response. It employs a ferroelectrically modulated polarization-sensitive photoelectric synapse device, comprising a ferroelectric material layer, a polarization-sensitive semiconductor channel layer, and corresponding gate and electrode structures. The ferroelectric material layer is introduced into the transistor structure as a gate dielectric to form a tunable ferroelectric polarization field under an applied electric field. The polarization-sensitive semiconductor material exhibits anisotropic photoelectric response characteristics to the polarization state of incident light. The ferroelectric polarization field achieves synaptic plasticity by modulating the channel carrier trapping and accumulation mechanism, enabling a nonlinear response to incident light with different polarization states, thereby obtaining a polarization-dependent photoelectric conversion response at the device level. Ferroelectric polarization modulation of the two-dimensional material introduces a controllable slow relaxation nonlinear process, giving the device output signal a memory effect. This allows for distinguishable dynamic response behavior in terms of time evolution and response amplitude characteristics under different polarization conditions. The aforementioned polarization-resolved nonlinear temporal response provides the physical basis for subsequent information encoding and differentiation, and storage cell calculation. Therefore, the key protection is given to the following: This invention uses a ferroelectrically regulated polarized light response device to obtain 16 distinguishable states to construct a storage pool.

[0037] The aforementioned storage pool system is used for identifying and distinguishing vehicle movement directions. Due to the introduction of polarized light, it is particularly suitable for low-visibility, degraded visual environments. By acquiring the time-series vectors corresponding to different vehicle movement directions at different polarization angles, the device generates a distinguishable dynamic response that can effectively identify the directionality of vehicle movement based on temporal characteristics.

[0038] Based on the aforementioned physical characteristics, this invention utilizes the ferroelectrically modulated polarization-sensitive photoelectric synapse to construct a device-level physical reservoir, achieving high-dimensional dynamic mapping of continuous optical inputs without relying on complex external storage or computing units. This physical reservoir can respond to the input timing pattern at the device level through ferroelectric polarization and carrier transport processes, causing the time-series optical signal to evolve into a discriminative dynamic conductance state in the device response. Furthermore, it retains historical input information at the device level, thereby meeting the requirements of recognition tasks for nonlinear response and time-series-related memory behavior.

[0039] Most importantly, this invention discovers that the polarization angle of illumination has a significant modulating effect on the state discrimination of the physical reservoir. At a 90° polarization angle, the device produces a stronger photoresponse amplitude and a more significant memory retention effect, resulting in a more distinct state distribution in the physical reservoir state space for different time-series inputs. However, at a 0° polarization angle, the device response tends to be linear, and the state distributions corresponding to different input sequences overlap, thereby reducing the discrimination capability. Figure 3 , Figure 4As shown, in the 10 sets of data tested, the light response basically did not cross under different polarization light input sequences at 90°, but the crossover was obvious at 0°.

[0040] Example 1 The present invention provides a method for calculating the storage pool and recognizing motion direction based on ferroelectric polarization response, which integrates a photosynaptic ferroelectric transistor into a polarization-sensitive neuromorphic vision system, the structure of which is as follows: Figure 1 As shown, the ferroelectric-modulated polarization-sensitive photoelectric synapse device includes the following structure: substrate 101, chromium-gold composite electrode 102, ferroelectric functional layer 103, and two-dimensional semiconductor channel layer 104. The substrate 101 is a silicon substrate with a silicon dioxide layer, or a silicon substrate covered with a silicon dioxide layer; Gold electrode 105, located above the silicon dioxide layer, serves as a bottom gate electrode for applying a control voltage to polarize the ferroelectric functional layer. The ferroelectric functional layer 103, covering the gold electrode and part of the substrate, is a polyvinylidene fluoride ferroelectric polymer film; A two-dimensional semiconductor channel layer 104 is placed on the ferroelectric functional layer and is an anisotropic palladium diselenide channel layer with a thickness of 40 nanometers. A chromium-gold composite electrode 102 forms ohmic contacts with both ends of the two-dimensional semiconductor channel layer. The source and drain electrodes are in contact with the underlying ferroelectric thin film functional chromium layer. The lower layer of the chromium-gold composite electrode is a chromium layer with a thickness of 10 nm, and the upper layer is a gold layer with a thickness of 50 nm. The ferroelectric functional layer 103 is a polyvinylidene fluoride ferroelectric polymer film, abbreviated as P(VDF-TrFE). The two-dimensional semiconductor channel layer 104 is an anisotropic palladium diselenide channel layer with a thickness of 40 nm. A gate voltage greater than its coercive field is applied to the organic ferroelectric polymer polyvinylidene fluoride P(VDF-TrFE) on the metal gate electrode using a probe station, ensuring that the P(VDF-TrFE) beneath the palladium diselenide in the two-dimensional material is fully polarized. Electrons or holes are injected into the two-dimensional material, thereby achieving the control of carrier concentration and photosynaptic characteristics.

[0041] The ferroelectric-modulated polarization-sensitive photoelectric synapse device of the present invention achieves long-term regulation of carrier concentration through the non-volatile polarization of the ferroelectric functional layer 103, endowing the sensor with photoelectric memory characteristics. At the same time, by utilizing the selective absorption of different polarized light by the palladium diselenide two-dimensional semiconductor channel layer 104, the polarization information is directly encoded as a difference in conductivity. The two work together to enable the device to generate a dynamic response to polarized light stimulation that depends on its own electrical characteristics. Thus, at the hardware level, the device realizes the perception, memory and primary spatiotemporal fusion processing of polarization visual information, providing a key basic device for building a high-performance, low-power biomimetic polarization vision system.

[0042] The fabrication of the photosynthetic ferroelectric transistor includes the following steps: 1) using ultraviolet lithography combined with electron beam evaporation to prepare a gold electrode 105 on a substrate 101; 2) using spin coating to prepare a P(VDF-TrFE) ferroelectric functional layer 103, and annealing it at 135°C for 4 hours to ensure the crystallization characteristics of the ferroelectric functional layer; 3) using a mechanical exfoliation transfer method to transfer a palladium diselenide two-dimensional semiconductor channel layer 104 to the surface of the ferroelectric functional layer 103; 4) using an in-situ transfer technique for chromium-gold composite electrodes 102 to transfer a pair of chromium-gold composite electrodes 102 pre-prepared by electron beam evaporation to the surface of the palladium diselenide two-dimensional semiconductor channel layer 104. The polarization-tunable ferroelectric functional film serves as a presynapse, and the channel layer can be modulated to generate changes in carrier concentration for the postsynapse. The weights between the presynaptic and postsynaptic neurons are updated and adjusted under electrical and optical stimulation, enabling the processing of polarization-related visual information. To achieve precise and continuous control of the polarization state of the incident laser, a combined optical path of half-wave plate 202 and polarizer 201 is used. This not only achieves fine and continuous control of the polarization angle but also simplifies operation, facilitating dynamic or precise polarization modulation of the incident light from the light source 300. The polarizer 201 is fixed in place, and the polarization direction output by its transmission axis system only allows polarized light components parallel to the transmission axis to pass through.

[0043] The half-wave plate 202, serving as a continuous polarization angle tuning unit, is placed before the polarizer. When linearly polarized light passes through the half-wave plate, its polarization direction rotates by an angle twice the angle between the incident polarization direction and the fast axis of the half-wave plate. Therefore, by rotating the half-wave plate, the polarization direction incident on the polarizer can be conveniently and continuously adjusted, thereby controlling the angle of the polarized light ultimately output by the system.

[0044] This invention provides a method for calculating the storage pool and identifying the direction of motion based on ferroelectric polarization response, which is used to realize the perception and classification of time-series moving targets, and is especially suitable for intelligent vision applications in low visibility or complex optical environments.

[0045] The PdSe2 / P(VDF-TrFE) ferroelectric photoelectric synapse device used in this invention exhibits a gradually increasing response with the number of light pulses under ferroelectric polarization modulation. This characteristic stems from the synaptic plasticity introduced by the ferroelectric field-assisted carrier trapping and accumulation mechanism, enabling the device to exhibit nonlinear response characteristics under polarized light stimulation and a memory effect that slowly decays and stabilizes in an intermediate state after the polarized light is removed. Figure 2 As shown.

[0046] The intensity and wavelength of the polarized light remain unchanged (520nm, 0.25W / cm). 2This invention employs 16 different 4-bit polarized light pulse time sequences: (1111 0111 1011 1101 1110 0011 0101 1001 0110 1010 1100 00010010 0100 1000 0000), where 0 represents no polarized light pulse and 1 represents a polarized light pulse. Due to the nonlinear response of the device to polarized light stimulation, the PdSe2 / P(VDF-TrFE) ferroelectric photoelectric synapse device responds differently to these 16 light pulse time sequences, and these responses can be effectively distinguished by the light response results. To verify the ability of the physical reservoir constructed by the device to distinguish time-series inputs, 4-bit time-series light pulses are used as input signals, with differences in the time order and combination of the different pulse sequences. Experimental results show that, under the same polarized light intensity, different time-series inputs can evolve into distinct final-state current values ​​in the device, and these final states are highly sensitive to the order of the input pulses. The results are shown below. A total of 10 groups were tested. Figures 3-4 As shown.

[0047] Based on the aforementioned physical characteristics, this invention utilizes the ferroelectrically modulated polarization-sensitive photoelectric synapse device to construct a device-level reservoir, enabling high-dimensional dynamic mapping of continuous optical inputs without relying on complex external storage or computing units. This physical reservoir can retain historical input information at the device level and naturally evolve time-series optical signals into discriminative dynamic states, thereby meeting the basic requirements of time-series recognition tasks for nonlinear response and relaxation memory characteristics.

[0048] Most importantly, this invention discovers that the polarization angle of illumination has a significant modulating effect on the state discrimination of the physical reservoir. At a polarization angle of 90°, the device produces a stronger light response amplitude and a more significant memory retention effect, resulting in a more distinct state distribution in the physical reservoir state space for different time-series inputs. However, at a polarization angle of 0°, the device response tends to be linear, and the state distributions corresponding to different input sequences overlap, thereby reducing the discrimination capability. Figures 7-8 As shown in the 10 sets of data tested, the light response basically did not cross under different polarization light input sequences at 90°, but the crossover was obvious at 0°.

[0049] Based on the aforementioned polarization-dependent characteristics, this invention utilizes ferroelectrically modulated polarization-selective light response to construct a polarization-tunable reservoir sensing mechanism, significantly enhancing the recognition performance of time-related features such as vehicle motion direction. By discretizing the vehicle motion process into multi-time-sequence inputs, the physical reservoir can effectively distinguish the temporal evolution patterns corresponding to different motion trajectories, achieving reliable identification of multiple vehicle motion directions.

[0050] The reservoir constructed from the described ferroelectrically modulated polarization-sensitive photoelectric synapse device (hereinafter referred to as the reservoir) is applied to a vehicle motion direction recognition scenario. Continuous visual stimulus signals act directly on the ferroelectric photoelectric synapse device in an optical form, causing the device to generate differentiated photoresponse currents under polarized illumination conditions. The sequence of these response currents serves as reservoir nodes, characterizing the temporal evolution of continuous images during vehicle motion. This approach avoids traditional image sensing, analog-to-digital conversion, and frame-by-frame digital processing, thereby significantly reducing system latency and energy consumption.

[0051] Based on the construction of a reservoir using a ferroelectrically modulated polarization-sensitive device, the time-series input bypasses traditional image sensing and analog-to-digital conversion processing, instead being directly input as optical pulses to the PdSe2 / P(VDF-TrFE) ferroelectrically modulated polarization-sensitive photosynaptic device. Under the influence of ferroelectric polarization modulation, the device simultaneously exhibits the following during optical stimulation: 1) a photoresponse based on changes in polarization angle; 2) nonlinear current modulation caused by carrier capture and release; and 3) relaxation response behavior to historical optical stimuli.

[0052] When different 4-bit time-series optical pulses are applied to the device, the device's output current evolves over time and eventually converges to a stable state related to the input sequence. This stable current value serves as the physical reservoir state output for the corresponding time sequence. Since the device's response to optical pulses depends not only on the number of pulses but also on their temporal order, different time-series inputs naturally evolve into distinct physical reservoir states within the device, thus achieving a high-dimensional mapping of temporal information at the device's physical level. Under 90° illumination conditions, where the device's polarization response is strongest, the device exhibits a stronger nonlinear response, resulting in a higher degree of differentiation in the physical reservoir state distribution corresponding to different time-series inputs.

[0053] Vehicle movement directions are categorized based on the initial direction of entry into the road and its subsequent trajectory, with eight categories defined: For vehicles entering the road from the right: straight (RS), left turn (RL), right turn (RR), and U-turn (RT); For vehicles entering the road from the left: straight (LS), left turn (LL), right turn (LR), and U-turn (LT). These eight movement directions exhibit different temporal evolution path characteristics from a top-down perspective and are the target categories for subsequent temporal sequence identification.

[0054] Regarding dataset generation: Time-series image data for simulating vehicle motion under real road conditions is generated through rule-driven traffic scene simulation to achieve systematic training and performance evaluation of the reservoir computing model. The dataset is generated using a rule-based traffic simulation program implemented in C++, which comprehensively incorporates road geometric constraints, multi-lane traffic organization structures, motion direction elements, and random noise disturbances to simulate degraded visual scenes under complex environmental conditions such as low visibility.

[0055] The simulated image data is rendered based on a 36×36 pixel grayscale image grid. The image uses a top-down view to construct a unified and structurally consistent road layout. The road network consists of a pair of mutually orthogonal corridor-like roads, each with a width or height of 12 pixels, centered along its corresponding coordinate axis. The roads are further divided into multiple sub-lanes to support multi-directional vehicle traffic, introducing spatial diversity while maintaining overall geometric consistency. Road boundaries and lane dividers are encoded using stable grayscale intensity values, serving as reference markers for subsequent motion recognition and spatial feature extraction, thereby enhancing the recognizability and consistency of the spatial structure in the dataset.

[0056] In the simulation, each vehicle target is represented by a 3×3 pixel high grayscale block to form clear and compact moving target features. All vehicle motion sequences are initialized from a designated sub-lane in the left-hand area of ​​the horizontal road to unify initial conditions; at the same time, controllable variations are allowed within the range of lane selection and spatial offset, thus balancing standardization and diversity. Each data sample consists of four consecutive frames to characterize the temporal evolution of vehicle motion.

[0057] The dataset contains eight categories of vehicle movement directions, defined based on real-world driving behavior from a top-down perspective. These directions include straight-ahead, left-turn, right-turn, and U-turn trajectories, as well as corresponding mirrored movement patterns depending on the side the vehicle enters the road. The generation of vehicle trajectories follows preset kinematic rules to ensure continuous lane movement and geometric consistency. Straight-ahead trajectories advance laterally along the horizontal lane; turning trajectories utilize controlled lane changes to allow vehicles to move from horizontal to vertical or opposite lanes. These rules ensure significant spatial differentiation between different movement directions while maintaining physical consistency. In one specific configuration, the simulation system generates 600 sets of sample data, with each movement direction consisting of four consecutive images, resulting in a total of 4,800 grayscale images. All images maintain consistency in spatial dimensions and grayscale encoding rules to ensure the overall standardization and comparability of the dataset.

[0058] To simulate visual degradation effects in low-visibility environments, a pixel-level noise model is introduced into the generated grayscale image. This noise model includes sparse, bright noise points and low-intensity background perturbations to simulate brightness anomalies caused by light scattering and random interference at the sensor level. This degradation processing improves the applicability of the dataset for evaluating the robustness of visual perception models under complex environmental conditions. Through the above dataset construction method, this invention can generate temporal motion image data with consistent structure, clear categories, and environmental degradation characteristics under controllable conditions, providing reliable data support for reservoir computing and the training and testing of related visual recognition models.

[0059] These parts are fed into the neural network for training, and some are used for testing.

[0060] For Reservoir Computing (RC), during the data preprocessing stage, the time-series dynamic images (the four images above) are converted into 4-bit input sequences, resulting in a total of 1296 sets of input data.

[0061] like Figure 5 As shown, for each 4-bit input sequence, one of the 10 photosynaptic current response values ​​obtained under the same input conditions is randomly selected as the output label value corresponding to that input sequence, thus constructing a data sample set for network training. The neural network is then fed the results of this dataset along with the corresponding vehicle direction. After training, a data sample with an unknown motion direction is input into the neural network, which identifies the vehicle direction of the response. The motion direction corresponding to the output node with the highest probability or confidence value is taken as the determination result of the vehicle's current motion direction.

[0062] The processed time-series dynamic data is then classified and identified using software. The identification process is based on a fully connected neural network model, such as... Figure 6 As shown, the neural network includes an input layer, at least one hidden layer, and an output layer. Its network structure is 1296×128×8, and it is built and implemented using the Python programming language in the TensorFlow and Keras deep learning framework.

[0063] In the fully connected neural network of this invention, each input node of the input layer (1296 input nodes) is used to receive a state feature value output by the physical reservoir; the state feature value is the stable current obtained after the time-series optical pulse signal acts on the ferroelectric photoelectric synapse device; the input layer is used to input only the high-dimensional time-series feature vector processed by the physical reservoir into the subsequent structure. In this way, the input layer can completely retain the nonlinear mapping result of the physical reservoir on the vehicle motion time evolution information, providing a high-dimensional feature basis for subsequent discrimination.

[0064] The hidden layer comprises 128 hidden nodes and is connected to the input layer via a fully connected manner. It performs weighted summation and nonlinear mapping on the 1296-dimensional reservoir state features provided by the input layer; it compresses and reassembles the high-dimensional, redundant reservoir output features; and it extracts a discriminative feature subspace representation related to vehicle motion direction determination. Here, it is assumed that this network has already been trained. The hidden layer, through the correspondence between the trained weight parameters and directions, distinguishes feature patterns corresponding to different motion directions, thereby enhancing the accuracy of motion direction recognition in subsequent output layers. The output layer includes 8 output nodes and 8 directions. Each output node corresponds to a preset vehicle motion direction category. Based on the results of the hidden layer, the output layer calculates the probability value or confidence value corresponding to each motion direction. The motion direction corresponding to the output node with the largest probability value or confidence value is used as the judgment result of the vehicle's current motion direction.

[0065] With the above settings, the output layer can classify and identify various vehicle movement directions.

[0066] Through the above technical solution, the storage pool calculation and motion direction recognition method based on ferroelectric polarization response of the present invention has achieved significantly improved recognition performance in vehicle motion direction recognition applications.

[0067] In one specific embodiment, the reservoir system was trained and tested under different light polarization states. The results showed that when the incident light polarization direction was 90°, the system exhibited a faster convergence speed during training and achieved a significantly improved motion direction recognition accuracy after training was completed. Specifically, under the optimal polarization response illumination condition, the system achieved a recognition accuracy of 96.3% for eight types of vehicle motion directions; while under the weaker polarization response illumination condition of 0°, the recognition accuracy was only 76.7%. Figure 7 As shown.

[0068] The corresponding confusion matrix further reflects the recognition results of the method of the present invention across eight motion direction categories. Under both polarization states, the recognition results for each category are mainly distributed along the diagonal of the confusion matrix, indicating that the method can effectively distinguish and stably identify multiple motion directions. Compared to 0° polarized illumination, under 90° polarized illumination, the concentration of diagonal elements is significantly improved, and misjudgments at off-diagonal positions are significantly reduced. This demonstrates that the present invention can effectively reduce the confusion between different motion directions under the stated polarization state, thereby improving the ability to distinguish and recognize time-evolving motion features.

[0069] To evaluate the separability of eight vehicle motion directions under different polarization states, a low-dimensional spatial mapping analysis was performed on the high-dimensional feature states obtained by the system to observe the distribution of different motion directions in the feature space. This mapping process projects the high-dimensional feature states into a two-dimensional space while preserving local adjacency relationships in the feature space, thereby reflecting the clustering characteristics and geometric distribution relationships of different motion direction categories in the feature space. In the mapping results, different vehicle motion directions form corresponding feature distribution regions. Under 90° polarized illumination conditions, such as... Figure 8 As shown, the characteristic states corresponding to each direction of motion exhibit a well-defined and mutually separated clustered distribution in two-dimensional space; while under 0° polarized illumination conditions, as... Figure 9 As shown, the feature distribution regions corresponding to different motion directions exhibit significant diffusion and partial overlap. These results further demonstrate that polarization modulation can effectively enhance the separability between the output states of the physical reservoir, thereby improving the ability to distinguish and identify the vehicle's motion direction.

[0070] like Figures 10-11 The experimental results confirm that the polarization-resolved optoelectronic device based on ferroelectric polarization modulation effectively acquires the temporal evolution characteristics of motion information sensing by controllably selecting and enhancing the polarization state of incident light. This ferroelectric polarization-modulated sensing mechanism can significantly suppress the interference of non-directional scattered light on the signal under complex imaging conditions such as visual degradation scenes in strongly scattering media, avoiding the collapse and aliasing of motion features in the feature space. This significantly improves the geometric separability of motion-coded features and ultimately achieves highly reliable differentiation and recognition of multi-directional motion information.

[0071] As can be seen, on a degraded image dataset incorporating simulated atmospheric scattering effects, the recognition performance of this system under 90° polarization is superior to that under 0° polarization and traditional intensity imaging methods. This confirms that the present invention, by utilizing polarization dimension information, can effectively suppress scattered light interference, extract motion features from degraded images, and achieve anti-interference capabilities.

[0072] The above specific embodiments are used to explain and illustrate the present invention, and are only preferred embodiments of the present invention, not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for calculating and identifying the motion direction of a storage cell based on ferroelectric polarization response, characterized in that, Includes the following steps: S1, the time-series dynamic image data of the moving target to be identified is encoded into a multi-bit time-series light pulse signal, and the time-series dynamic image data of the moving target to be identified is a continuous multi-frame image; S2, The time-series optical pulse signal generated in step S1 is input into the ferroelectrically modulated polarization-sensitive photoelectric synapse device for stimulation; Under the time-series optical stimulation, the photocurrent output by the polarization-sensitive photoelectric synapse device exhibits nonlinear dynamic evolution and reaches a steady-state response with the input of the pulse sequence; The steady-state photocurrent or equivalent conductance value after the evolution is completed will be used as the state output of the physical reservoir. S3, input the state output of the physical reserve pool described in step S2 into the subsequent fully connected neural network to complete the classification and recognition of the motion pattern of the moving target; In step S1, the moving target is a vehicle, the time-series dynamic image data is a 36×36 pixel grayscale image from a top-down perspective, and each data sample consists of four consecutive frames of images; the multi-bit time-series optical pulse signal is a 4-bit signal, where "1" represents an excitation optical pulse with a specific polarization direction, and "0" represents an excitation pulse with no light or orthogonal polarization direction. In step S2, the ferroelectrically modulated polarization-sensitive photoelectric synapse device includes a polarization light-sensitive unit, which includes at least one photosynaptic ferroelectric transistor. The two-dimensional semiconductor channel layer of the photosynaptic ferroelectric transistor is composed of a two-dimensional semiconductor material with optical anisotropy, and its ferroelectric functional layer is composed of a ferroelectric material. A polarization light modulation module is disposed on the light receiving path of the photosynaptic ferroelectric transistor to provide the photosynaptic ferroelectric transistor with an adjustable polarization direction; The ferroelectric functional layer serves as a presynapse, and the two-dimensional semiconductor channel layer serves as a postsynapse. Under the regulation of ferroelectric polarization, the device exhibits enhanced nonlinear response characteristics to illumination conditions corresponding to the direction of strongest polarization response, thereby achieving deep integration of polarization information and nonlinear synaptic dynamics. In step S3, the time-series optical pulse signal under the motion direction encoded in step S1 is matched and labeled with the physical reservoir state output obtained in step S2 to construct a supervised learning dataset for training the readout layer neural network, specifically including: S301, Encode each frame of moving target time-series dynamic image data with known motion direction into a multi-bit time-series light pulse signal to obtain multiple sets of input sequences; S302, each set of input sequences is input to the ferroelectric photoelectric synapse device, repeated measurements are performed, and one of the multiple photocurrent response final state values ​​is selected as the physical reservoir state output value corresponding to the input sequence; S303, construct a set of state vectors from the physical reservoir state output values ​​corresponding to the multiple sets of input sequences, and associate them with the corresponding vehicle motion direction labels, thereby forming a supervised learning dataset for training the readout layer neural network, wherein the readout layer neural network is a fully connected neural network.

2. The method for calculating and identifying the motion direction of a storage pool based on ferroelectric polarization response as described in claim 1, characterized in that... In step S1, the moving target temporal dynamic image data includes temporal data samples consisting of eight different motion directions and multiple consecutive grayscale images. The eight different directions of movement include: going straight, turning left, turning right, and making a U-turn when entering from the right side of the road, and going straight, turning left, turning right, and making a U-turn when entering from the left side of the road.

3. The method for calculating and identifying the motion direction of a storage pool based on ferroelectric polarization response as described in claim 1, characterized in that: In step S2, while performing light stimulation, an initialization voltage pulse with an absolute value greater than the coercive field of the ferroelectric thin film is applied to the gate of the ferroelectric modulated polarization-sensitive photoelectric synapse device to regulate the channel and interface state. After each timing signal processing, a reverse gate voltage is applied to erase the ferroelectric polarization state, ensuring that the device starts in the same state before the next timing signal processing.

4. The method for calculating and identifying the motion direction of a storage pool based on ferroelectric polarization response as described in claim 1, characterized in that, The photosynthetic ferroelectric transistor includes: The substrate is a silicon substrate covered with a silicon dioxide layer; The gold electrode is located on the silicon dioxide layer; The ferroelectric functional layer, covering the gold electrode and part of the substrate, is a polyvinylidene fluoride ferroelectric polymer film; A two-dimensional semiconductor channel layer, placed on the ferroelectric functional layer, is an anisotropic palladium diselenide channel layer with a thickness of 40 nanometers. The chromium-gold composite electrode forms ohmic contacts with both ends of the two-dimensional semiconductor channel layer. The lower layer of the chromium-gold composite electrode is a chromium layer, and the upper layer is a gold layer. The polarization light modulation module irradiates the photosynaptic ferroelectric transistor, providing polarized light stimulation to the two-dimensional semiconductor channel layer. By applying a gate voltage higher than the coercive field of the ferroelectric material on the chromium-gold composite electrode, the polyvinylidene fluoride ferroelectric polymer film is fully polarized, thereby injecting electrons or holes into the palladium diselenide channel layer, achieving regulation of carrier concentration and photosynaptic characteristics.

5. The method for calculating and identifying the motion direction of a storage pool based on ferroelectric polarization response as described in claim 1, characterized in that, In step S2, by adjusting the polarization direction of the incident light, the state space distribution of the physical reservoir is actively controlled. Under the 90° polarization direction, the ferroelectric control polarization-sensitive photoelectric synapse device enables the final state current corresponding to different input sequences to form a high separation distribution in the state space, thereby achieving its high classification performance in the time sequence recognition task.

6. The method for calculating and identifying the motion direction of a storage pool based on ferroelectric polarization response as described in claim 1, characterized in that, The readout layer neural network includes: The input layer, whose number of nodes is the same as the dimension of the state vector, is used to receive the state vector composed of the state output values ​​of the physical reserve pool. At least one hidden layer is used for nonlinear transformation and feature extraction of the state vector; The output layer, with the same number of nodes as the number of motion direction categories to be classified, is used to output the classification results.

7. The method for calculating and identifying the motion direction of a storage pool based on ferroelectric polarization response as described in claim 6, characterized in that, Using a trained fully connected neural network, the time-series dynamic image data of the moving target in the direction of motion to be identified in step S1 is converted into a 4-bit time-series light pulse signal, generating a total of 1296 different input sequences. In step S3, the input layer contains 1296 neurons, which correspond one-to-one with the state output dimension of the physical reservoir, and fully receive the 1296-dimensional state feature vector output from step S2. The hidden layer contains 128 neurons, and the output layer contains 8 neurons, which correspond to 8 preset vehicle motion directions, output the probability value of each direction category, and take the category with the highest probability value as the final recognition result.

Images