Motion intention pre-judgment method of dynamic fuzzy trajectory

Abstract

The invention discloses a motion intention pre-judgment method for a dynamic fuzzy trajectory, and relates to the technical field of automatic driving, and the method comprises sensor data collection, multi-source data fusion and motion intention decision-making, and specifically comprises the steps: building a physical field induction network, integrating a magnetoresistive sensor array and an acoustic resonance sensor through a composite sensor unit, and obtaining a motion intention pre-judgment result; carrying out low-delay acquisition and transmission on a physical field signal by combining spectrum modulation of a vehicle-end composite antenna; according to the invention, low-delay data acquisition is realized through a composite sensor and a collaborative acquisition network, the data processing efficiency is improved by combining a distributed database and a three-layer reasoning network, parameters are optimized by utilizing edge-cloud collaborative feedback and a hybrid algorithm, and the problems of high delay, poor adaptability and the like in the traditional technology are solved. The detection precision is improved, millisecond-level response is inquired, the pre-judgment accuracy is improved, and the automatic driving safety is remarkably enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to a motion intention prediction method for dynamic fuzzy trajectory. BACKGROUND

[0002] In recent years, automatic driving technology has made significant progress, and is gradually moving from research and development testing to commercial application. From L2 level assisted driving to higher level L3, L4 level automatic driving, the perception and decision-making ability of vehicles in complex traffic environment is continuously improved. However, to achieve completely reliable automatic driving, there are still many technical challenges, among which the accurate prediction of the motion intention of traffic participants is crucial. In complex traffic scenarios, the motion trajectories of vehicles, pedestrians, non-motor vehicles and other traffic participants show high dynamicity and uncertainty, which puts high requirements on the perception, decision-making and control of automatic driving systems. Accurate prediction of the motion intention under these dynamic fuzzy trajectories can plan the vehicle driving path in advance, optimize the driving decision, and significantly improve the safety and traffic efficiency of automatic driving.

[0003] Traditional motion intention prediction technology mainly relies on sensor data acquisition, multi-source data fusion and complex decision-making algorithms to infer the intention of traffic participants. In the sensor data acquisition link, although laser radar, camera, millimeter wave radar and other sensors can obtain rich environmental information, they are limited by hardware performance and signal transmission speed, and there is a certain delay in data acquisition. For example, it may take tens of milliseconds for a common laser radar to complete an environmental scan and transmit data to the perception module. The multi-source data fusion process is also time-consuming, and the format, frequency and accuracy of different types of sensor data are different, and there are many technical difficulties in efficiently integrating and extracting effective information. In terms of decision-making algorithms, in order to ensure the accuracy of prediction, complex machine learning or deep learning models are often used, such as models based on convolutional neural network (CNN) and recurrent neural network (RNN). However, these models have huge computational load, and when running on a vehicle-mounted computing platform, they often produce a delay of 100-500 ms, resulting in a lag in motion intention prediction.

[0004] This lag brings serious hidden dangers in actual driving scenarios. In highway scenarios, when encountering a sudden deceleration or lane change of the vehicle in front, due to the prediction lag, the autonomous vehicle may not be able to respond in time, increasing the risk of rear-end collisions. In intersection scenarios, the delay in predicting the intentions of vehicles crossing or pedestrians may cause the autonomous vehicle to miss the best passing opportunity, leading to collision accidents or traffic congestion. In complex urban scenarios, the dynamic and ambiguous trajectories of multiple traffic participants are intertwined, and the prediction lag is more likely to cause the autonomous vehicle to be in a decision-making dilemma, reducing the passing efficiency. The traditional technical route has been committed to shortening this delay by improving the sensor refresh rate, optimizing the neural network architecture, and using edge computing, but the effect is limited, and the cost is high, which is difficult to fundamentally solve the problem.

[0005] In view of this, a motion intention prediction method for dynamic ambiguous trajectories is provided to overcome the above problems. SUMMARY

[0006] The purpose of the present application is to provide a motion intention prediction method for dynamic ambiguous trajectories to solve the problems raised in the background art.

[0007] To solve the above technical problems, the motion intention prediction method for dynamic ambiguous trajectories provided by the present application includes sensor data acquisition, multi-source data fusion and motion intention decision, specifically: A physical field sensing network is built, a composite sensor unit is used to integrate a magnetoresistance sensor array and a sound wave resonance sensor, and a vehicle-end composite antenna is used for spectrum modulation to realize low-delay acquisition and transmission of physical field signals; A biological behavior pattern database is constructed, traffic participant behavior data is obtained through a network of coordinated acquisition of heaven, earth and vehicles, and a micro-expression recognition technology is introduced to establish a probability mapping model of emotions and behavior patterns; A three-level probability reasoning network is constructed, and the fusion decision of physical field features, behavior patterns and environmental information is made through a convolutional autoencoder, a Bayesian network and a D-S evidence theory; An edge-cloud collaborative feedback mechanism is established, and the dynamic optimization of sensor parameters and model parameters is realized through an adaptive PID algorithm and a deep Q network.

[0008] Further, the composite sensor unit integrates the magnetoresistance sensor and the sound wave resonance sensor in a buried modular device, wherein the magnetoresistance sensor uses a three-axis magnetoresistance chip and is equipped with a temperature compensation circuit.

[0009] Further, the vehicle-end composite antenna uses a paper folding structure flexible circuit board, integrates a dual-channel spectrum modulator, independently processes electromagnetic signals and sound wave signals through orthogonal frequency division multiplexing, and enhances the target echo signal in combination with a pulse code modulation generator.

[0010] Further, the sky-ground vehicle cooperative collection network includes a UAV group, a vehicle-mounted camera and a distributed cooperative positioning technology, wherein the UAV performs 120-degree wide-angle video collection at a height of 100 meters and a speed of 15 km / h, the vehicle-mounted camera is configured with a dynamic aperture and an optical anti-shake system, and a behavior feature is extracted by combining a space-time attention mechanism.

[0011] Further, the three-level probability inference network includes: A physical field feature extraction layer adopts a convolutional autoencoder to reduce dimension and denoise the magnetic resistance and sound wave signals; A behavior pattern matching layer calculates a probability distribution of a motion intention through a Bayesian network; A dynamic decision layer corrects the intention probability based on a D-S evidence theory and fuses environmental information.

[0012] Further, the biological behavior pattern database adopts a distributed heterogeneous architecture, a blockchain layer ensures data unforgeability through a practical Byzantine fault tolerance algorithm, a graph database layer constructs a relationship network including behavior nodes, environmental factor nodes and behavior motivation nodes, and an index is dynamically adjusted by combining a reinforcement learning optimization module.

[0013] Further, the edge-cloud cooperative feedback mechanism adopts a double closed-loop control: An inner loop dynamically adjusts sensor parameters through an adaptive PID algorithm, and PID parameters are jointly controlled based on an error gradient, an accumulated amount and an acceleration; An outer loop optimizes model parameters through a deep Q network combined with a dynamic reward function, and safety, accuracy and efficiency are taken as three-dimensional reward dimensions.

[0014] Further, the feedback optimization mechanism introduces a hybrid intelligent algorithm of double-population particle swarm optimization and differential evolution algorithm, combines a chaotic mapping initialization and a Cauchy mutation strategy, and balances physical field perception accuracy, behavior matching accuracy and computing resource consumption through a multi-objective optimization function.

[0015] Compared with the prior art, the present application has the following beneficial effects: 1. Sensor data collection optimization effect: The composite sensor unit integrates a magnetic resistance sensor array and a sound wave resonance sensor, the magnetic resistance sensor adopts a three-axis chip and a temperature compensation circuit, the detection accuracy reaches 0.1 nT and the temperature drift interference is eliminated, the sound wave resonance sensor realizes ±15° beam scanning through a 7*7 piezoelectric ceramic array, the ranging accuracy is ±5 cm within 100 meters, and the problem of unstable detection accuracy of a traditional sensor is solved.

[0016] Vehicle-end composite antenna: adopts a flexible circuit board with a folding structure, integrates a dual-channel spectrum modulator, realizes independent processing of electromagnetic and acoustic signals through OFDM technology, enhances echo signals by combining with a PCM generator, effectively improves detection distance in complex environments, and solves the problem of insufficient detection range of traditional sensors.

[0017] Sky-ground vehicle cooperative collection network: the unmanned aerial vehicle collects 120° wide-angle video at a height of 100 meters and a speed of 15 km / h, the vehicle-mounted camera is equipped with a dynamic aperture and an optical anti-shake system, combined with STAM to extract features and introduce micro-expression recognition to establish an emotion-behavior probability mapping model, filling the gap in behavior motivation data and solving the problem of limited perspective and failure in special scene prediction in traditional data collection.

[0018] 2. Advantages of biological behavior pattern database Distributed heterogeneous architecture: the blockchain layer ensures data tamper-proofing through the PBFT algorithm, the graph database layer constructs a relationship network containing behavior, environment, and motivation nodes, and the complex query response time is shortened to milliseconds, solving the problems of low security, complex relationship processing, and low query efficiency in traditional databases.

[0019] Reinforcement learning optimization module: through the cooperation of simulated annealing and genetic algorithm optimization, the index structure is dynamically adjusted to improve data query efficiency and provide efficient data support for real-time prediction.

[0020] 3. Effectiveness of motion intention prediction model Three-level probability inference network: the physical field feature extraction layer uses CAE for dimensionality reduction and denoising, the behavior pattern matching layer calculates the intention probability distribution through the Bayesian network, and the dynamic decision-making layer corrects the probability based on the D-S evidence theory fusion environment information, realizes multi-dimensional information deep fusion decision-making, predicts the intention in advance and outputs the probability distribution, provides quantitative basis for risk assessment.

[0021] Data dynamic analysis method: physical field data uses VMD and HHT combined analysis to extract 0.1Hz-100Hz characteristic frequency; biological behavior pattern uses cloud-Markov chain model for processing, describing the fuzziness and randomness of behavior; the vehicle-mounted computing platform uses heterogeneous multi-core architecture and task dynamic scheduling algorithm to improve computing resource utilization and reduce energy consumption, ensuring the real-time performance of complex scene prediction.

[0022] 4. Feedback and optimization mechanism Edge-cloud collaborative feedback: the edge node uses NVIDIA Jetson AGX Orin platform for real-time data preprocessing to reduce transmission delay; in the double-loop control, the inner loop uses adaptive PID algorithm to dynamically adjust sensor parameters, the accuracy of the magnetic resistance sensor is improved, and the sound wave ranging error is reduced; the outer loop optimizes model parameters through DQN combined with dynamic reward function, and the accuracy of complex scene prediction is improved.

[0023] Hybrid intelligent optimization algorithm: DPSO is combined with DE, chaos mapping initialization and Cauchy mutation strategy are introduced, the perception accuracy, matching accuracy and resource consumption are balanced through a multi-objective optimization function, the resource consumption is reduced, and the resource bottleneck of traditional algorithms is broken through.

[0024] Flink real-time computing framework: minute-level model parameter updating is realized, dynamic environment changes such as road construction and large-scale activities are quickly adapted, and the system adaptability is enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The principle diagram of the motion intention prediction method of the dynamic fuzzy trajectory. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0027] Please refer to Figure 1 The application provides a technical solution: Please refer to Figure 1 The embodiment of the motion intention prediction method of the dynamic fuzzy trajectory is shown: I. Physical field induction network construction: 1. Road infrastructure sensor: Traditional sensors have single function and need to be combined with multiple types of sensors, which not only increases the difficulty of data fusion, but also causes high data acquisition delay due to the limitation of hardware performance. The method uses a composite sensor unit, integrates a magnetoresistance sensor array and a sound wave resonance sensor in a buried modular device, and fundamentally changes the data acquisition mode.

[0028] Magnetoresistance sensor: a magnetoresistance sensor based on tunnel magnetoresistance effect has a 0.1nT magnetic field change detection accuracy, which is much higher than that of traditional magnetic sensors and can capture extremely weak geomagnetic field changes caused by the movement of vehicles and pedestrians carrying metal objects. The built-in three-axis magnetoresistance chip can detect the magnetic field changes in X, Y and Z axis directions at the same time, compared with single-axis or double-axis detection, it can more comprehensively and accurately obtain target motion trajectory information. Considering the influence of environmental temperature on the detection results of the magnetic sensor, the sensor is equipped with a temperature compensation circuit, which can monitor the environmental temperature in real time and correct the detection data according to the temperature-magnetoresistance characteristic curve, eliminate the interference of environmental temperature drift on the detection results, and solve the problem of unstable detection accuracy of traditional sensors in different temperature environments.

[0029] Acoustic resonance sensor: The acoustic resonance sensor using phased array technology realizes a beam scanning range of ±15° through a 7x7 piezoelectric ceramic array. Compared with the fixed beam direction of traditional acoustic sensors, it can realize multi-angle detection of the target and effectively expand the detection range. The time delay of the reflected wave is calculated using the cross-correlation algorithm, which can achieve a ranging accuracy of ±5 cm within a range of 100 meters. Compared with the traditional time-of-flight (ToF) method, the algorithm has stronger anti-interference ability and more accurate ranging in complex environments. The adaptive filtering module built into the sensor uses the recursive least squares (RLS) method to update the filtering parameters in real time, which can dynamically adapt to complex environmental clutter such as 50Hz power frequency interference and traffic noise. Compared with the traditional fixed parameter filtering method, it can more effectively suppress interference and ensure the purity of the detection signal.

[0030] Integrated modularization: The two sensors are integrated into a buried modular device, which reduces the road space occupation and facilitates large-scale deployment. On the other hand, the unified data interface and communication protocol simplify the data transmission process, reduce the difficulty of multi-source data fusion, and solve the problems of data transmission delay and complex fusion caused by traditional multi-sensor combination.

[0031] 2. Vehicle end receiving and enhancement device: The composite antenna designed with a flexible circuit board of origami structure realizes 180° folding through shape memory alloy driving. The antenna integrates a dual-channel spectrum modulator, which modulates electromagnetic signals (300MHz-3GHz frequency band) and acoustic signals (10kHz-50kHz frequency band) to different carriers respectively, and uses orthogonal frequency division multiplexing (OFDM) technology to avoid signal aliasing. Traditional multi-signal processing methods often use simple time division multiplexing or frequency division multiplexing, which can easily cause signal interference in complex electromagnetic environments. Through spectrum modulation and OFDM technology, efficient coexistence and independent processing of electromagnetic and acoustic signals are realized, improving the stability and reliability of signal transmission. At the same time, the vehicle-mounted pulse code modulation (PCM) generator can emit electromagnetic and acoustic pulses with specific coding sequences, and enhance the target echo signal through matched filtering technology. In complex environments, it can effectively distinguish target signals and interference signals, effectively improve the detection distance, and solve the problem of insufficient detection range of traditional sensors in complex environments.

[0032] II. Construction of biological behavior pattern database: 1. Behavior data collection and analysis: Traditional motion intention prediction technology relies on single vehicle-mounted sensor for data collection, which has limited data perspective and is difficult to capture complex behaviors of traffic participants. Moreover, it lacks consideration of psychological factors behind behaviors, resulting in poor generalization ability of prediction models. This method constructs a ground-air-vehicle collaborative collection network. UAVs collect 120° wide-angle video at a height of 100 meters and a speed of 15 km / h, and achieve a positioning accuracy of ±20 cm by combining distributed collaborative positioning technology.

[0033] The vehicle-mounted camera is equipped with a dynamic aperture and an optical image stabilization system, which can still output 1080P / 60fps high-definition video in low-light night environments, solving the problem of poor data quality of traditional sensors in complex lighting conditions. The data processing end uses a spatio-temporal attention mechanism (STAM) to extract spatial features and optical flow temporal features through a double-stream convolutional neural network (CNN), and combines an LSTM network to predict behavior trends.

[0034] Micro-expression recognition technology is introduced to establish a probability mapping model of emotional state and behavior pattern through facial key point detection and electromyographic signal analysis. In the field of autonomous driving, no one has ever incorporated human emotional factors into the motion intention prediction system, but in fact, emotions can significantly affect the behavior decisions of pedestrians and drivers. For example, a pedestrian in an anxious state may suddenly accelerate and cross the road. In this way, the data gap in the behavior motivation layer is filled, providing a new data dimension for motion intention prediction, making the prediction model more realistic, and effectively solving the problem of prediction failure in special scenarios caused by the lack of behavior motivation analysis in traditional technology.

[0035] 2. Database construction and optimization: Traditional database architecture has problems such as low data security, complex relationship processing, and low query efficiency when storing and querying traffic behavior data. This method uses a distributed heterogeneous database architecture, combining the advantages of blockchain and graph databases. The blockchain layer uses a consortium chain model to achieve data consensus between nodes through the practical Byzantine fault tolerance (PBFT) algorithm, ensuring data cannot be tampered with, which solves the problem of data tampering and poor security in traditional databases, providing a guarantee for the credibility of autonomous driving data and having important significance in scenarios such as autonomous driving responsibility identification.

[0036] The graph database layer uses an attribute graph model to construct a relationship network containing 12 types of behavior nodes, 8 types of environmental factor nodes, and 20 types of behavior motivation nodes, which can clearly present the complex relationships between traffic participants' behaviors, environments, and motivations. Compared to the table structure of traditional databases, the efficiency of complex relationship queries is greatly improved. For example, when querying "the influencing factors of pedestrian acceleration behavior on school road during the morning rush hour", traditional databases may need to perform complex association queries across multiple tables, while the graph database of this method can quickly obtain the result through the relationship network.

[0037] The reinforcement learning optimization module is introduced to realize dynamic adjustment of the database index structure through simulated annealing algorithm (SA) and genetic algorithm (GA) collaborative optimization, so as to shorten the complex query response time to milliseconds. The limitation of traditional static index of the database is broken, and the index can be automatically optimized according to the actual query demand and data change. In the automatic driving scene with extremely high real-time requirement of traffic data, the data query efficiency is greatly improved, and efficient data support is provided for real-time motion intention prediction. At the same time, this dynamic optimization mechanism can also continuously optimize the database performance with the growth of data volume and the change of query mode, which has a forward-looking advantage in coping with more complex traffic scenes and massive data accumulation in the future.

[0038] III. Real-time prediction of motion intention 1. Cross-analysis model Traditional motion intention prediction mainly relies on single data or simple multi-source data superposition analysis, which is difficult to handle the fuzziness and uncertainty in complex traffic scenes. This method constructs a three-level probability reasoning network from three dimensions of physical field, behavior pattern and environmental information, and realizes accurate prediction of motion intention.

[0039] The first layer is the physical field feature extraction layer. Considering that the electromagnetic and acoustic signals collected by the magnetic resistance sensor and the acoustic resonance sensor have high dimensionality and much noise, if traditional filtering and feature extraction methods are used, it is difficult to effectively extract key features and the calculation amount is large. Therefore, a convolutional autoencoder (CAE) is used to automatically extract local features of the signal through the convolutional layer, reduce the data dimension through the pooling layer, and then reconstruct the signal through the deconvolutional layer. In this way, the dimension is reduced and the noise is removed. Without manually designing complex feature extraction rules, the CAE can adaptively learn signal features, greatly reducing manual intervention and calculation amount.

[0040] The second layer is the behavior pattern matching layer. The behavior intention of traffic participants has probability and diversity, and the traditional matching method based on rules or simple statistics cannot accurately describe it. The Bayesian network is based on probability reasoning, which can combine a large amount of behavior pattern information in the biological behavior database with the current physical field characteristics to calculate the probability distribution of different motion intentions. It not only considers the prior probability of the behavior pattern, but also dynamically updates the probability according to real-time data, which is more consistent with the uncertainty of behavior in actual traffic scenes compared with traditional methods.

[0041] Third layer: dynamic decision layer: the actual traffic environment is complex and changeable, and environmental information such as weather and traffic events has an important influence on the motion intention, but traditional methods often ignore or simply process these information. Based on D-S evidence theory, real-time environmental information is taken as new evidence, and the intention probability obtained from the previous two layers is weighted and corrected by the evidence synthesis rule. It can effectively process uncertain information, fuse evidence from different sources, enhance the reliability and accuracy of decision-making, and provide a more comprehensive basis for motion intention prediction.

[0042] In addition, the system is built-in with a scene perception module, which takes into account the great differences in the behavior patterns of traffic participants in different special scenarios, and the traditional unified decision-making model cannot adapt. By automatically identifying 12 special scenarios and calling the exclusive decision-making model, it can optimize the decision-making for the characteristics of different scenarios and improve the accuracy and adaptability of the prediction.

[0043] It effectively solves the problems of difficult data feature extraction, inaccurate behavior intention matching and insufficient use of environmental information in traditional technology. It improves the efficiency of feature extraction of physical field signals and the accuracy of behavior intention matching.

[0044] And: it realizes the deep fusion decision of physical field signals, biological behavior patterns and environmental information, and can predict the motion intention of traffic participants in advance in complex traffic scenarios, providing more decision-making time for autonomous vehicles. At the same time, the model can also output the probability distribution of different intentions, providing quantitative basis for risk assessment of autonomous driving.

[0045] At the same time: it provides more accurate basic data for subsequent feedback optimization, so that the system can more accurately adjust the sensor parameters and model weights, further improving the prediction performance. For example, when the system finds that the prediction error is large in a certain type of scenario, it can optimize the parameters of the exclusive decision-making model for that scenario, forming a virtuous cycle.

[0046] 2. Dynamic analysis and processing of data: Physical field data has the characteristics of non-stationary and non-linear, and biological behavior patterns have uncertainty and dynamic changes, so traditional single analysis methods are difficult to effectively process these problems.

[0047] Physical field data analysis: for physical field data, a combination of variational mode decomposition (VMD) and Hilbert-Huang transform (HHT) analysis method is used. VMD can adaptively decompose complex physical field signals into multiple intrinsic mode functions, each of which represents a different characteristic component of the signal; HHT performs time-frequency analysis on each modal function to accurately extract 0.1Hz-100Hz characteristic frequency components. It can better handle non-stationary signals and uncover the hidden small changes in the signal, providing more abundant information for motion intention prediction.

[0048] Biological behavior pattern processing: For the uncertainty of biological behavior patterns, a cloud-Markov chain model is constructed. The cloud model converts qualitative behavior patterns into quantitative digital features (expectation, entropy, hyper-entropy), describing the fuzziness and randomness of behavior; Markov chain is used to predict the transition probability of behavior between different states. The combination of the two breaks the limitations of traditional simple classification or statistics of biological behavior, dynamically and probabilistically describes the changes of behavior patterns, and is more consistent with the complexity of behavior in actual traffic scenes.

[0049] Vehicle-mounted computing platform architecture: A heterogeneous multi-core architecture integrating FPGA, GPU and ARM processor is adopted, combined with a task dynamic scheduling algorithm, to solve the problem of limited computing resources of traditional vehicle-mounted computing platforms, which makes it difficult to efficiently process multiple types of data. FPGA is suitable for parallel processing of sensor data, GPU is good at processing images and complex algorithms, and ARM processor is used for system control and scheduling. The task dynamic scheduling algorithm dynamically allocates computing resources according to the characteristics and computing requirements of different tasks, greatly improving the utilization of computing resources.

[0050] Solved the problems of inaccurate traditional physical field data analysis, simple biological behavior pattern processing and waste of computing resources. The accuracy of physical field data feature extraction is improved, the reliability of biological behavior pattern prediction is improved, and the utilization of computing resources is improved.

[0051] Achieved deep dynamic analysis of physical field signals and biological behavior patterns, able to capture extremely small behavior changes and intention trends of traffic participants, such as predicting lane change intention when a vehicle makes a slight turning action. At the same time, the combination of heterogeneous multi-core architecture and dynamic scheduling algorithm enables the system to run complex algorithms with low power consumption, reducing the energy consumption of the vehicle-mounted computing platform.

[0052] At the same time: Provides protection for the stable operation of the system in extreme scenarios and complex tasks. Even in the case of a sharp increase in data volume or an increase in algorithm complexity, the real-time and accuracy of the prediction can be ensured through reasonable allocation of computing resources.

[0053] Four, intention prediction feedback and optimization: 1. Feedback mechanism: Traditional feedback systems mostly use single cloud processing or local independent optimization mode, which has the problems of high data transmission delay, untimely response, and inability to consider simultaneous optimization of sensor parameters and algorithm models. This method is based on the cross-layer collaborative idea of "edge intelligence front-end processing + cloud global optimization", and constructs an edge-cloud collaborative computing architecture.

[0054] The edge node selects the NVIDIA Jetson AGX Orin platform, which has a powerful computing power of 275 TOPS designed for edge computing scenarios, can directly preprocess 10 video streams and 50 sensor data in real time on the vehicle side, and can greatly reduce the bandwidth pressure and transmission delay of uploading data to the cloud. At the same time, the Kubernetes management cloud distributed computing cluster is adopted to realize the dynamic elastic allocation of computing resources, and the computing resources can be flexibly adjusted according to the traffic data volume changes in different periods and different regions, avoiding resource waste, and the resource utilization rate is improved compared with the traditional fixed resource allocation mode.

[0055] The feedback system adopts a double closed-loop control strategy. The inner loop adjusts the sensor parameters, and an adaptive PID algorithm is used to change the traditional PID parameter fixation and the difficulty of adapting to complex environmental changes. This algorithm dynamically adjusts the proportional (P), integral (I), and differential (D) parameters by real-time monitoring of the fluctuation amplitude of the magnetic field signal detected by the magnetic resistance sensor and the ranging error rate of the acoustic sensor. For example, near a substation with strong electromagnetic interference, the system automatically increases the differential parameter to speed up the suppression of interference signals, so that the sensor detection accuracy remains at ±0.05nT in a complex electromagnetic environment; in noisy construction sections, the adaptive adjustment of the integral parameter reduces the cumulative error of the acoustic signal due to environmental noise, and the ranging error is controlled within ±3cm, effectively solving the problem of inaccurate data caused by environmental interference of traditional sensors.

[0056] To solve the problem of insufficient anti-interference caused by fixed traditional PID parameters, an adaptive parameter updating model is proposed. The sensor error The dynamic relationship between the PID parameters is shown in the formula: Through the joint control of error gradient, cumulative amount and acceleration, real-time optimization of parameters is realized. For example, in an electromagnetic interference scene, the error change rate Drive increases, so that the detection accuracy of the magnetic resistance sensor is maintained at ±0.05nT (see formula one): Adaptive PID parameter updating model: (formula one); Where: , , are the proportional, integral, and differential parameters at time t; , , , are the initial parameters; , , are the adaptive coefficients; For sensor detection error (such as magnetic field deviation of magnetic resistance sensor or ranging error of acoustic wave sensor); For error rate of change, for dynamic adjustment To quickly suppress interference; For error accumulation, drive Eliminate static error; For error acceleration, optimization To deal with sudden interference.

[0057] Directly solve the problem of "environmental interference leading to inaccurate data" of traditional sensors, improve the accuracy of magnetic resistance sensors and reduce acoustic ranging errors through dynamic parameter adjustment, and reduce the delay source from the hardware perception layer.

[0058] The outer loop optimizes model parameters based on the deep Q network (DQN) algorithm of reinforcement learning, breaking through the limitations of traditional manual parameter adjustment or simple machine learning algorithm optimization. This algorithm uses actual driving data in the autonomous driving process (such as the actual driving trajectory after the vehicle makes a decision based on the predicted intention, and the actual interaction with surrounding traffic participants) as feedback signals, and through the construction of a "state-action-reward" model, the system learns through continuous trial and error. For example, when the system predicts that the vehicle in front has a lane change intention but it does not actually happen, a negative reward is given to prompt the model to adjust the parameters to reduce the misjudgment probability in this scenario; if the prediction is accurate and the vehicle avoids collision safely, a positive reward is given. In this way, the overall prediction accuracy of the model in complex traffic scenarios is improved, realizing the autonomous evolution of the algorithm model and solving the problem that traditional algorithms are difficult to adapt to dynamic traffic environments.

[0059] To realize the autonomous evolution of model parameters, this method designs a dynamic reward function (see Equation Two). With safety, accuracy, and efficiency as the three-dimensional reward dimensions, the model is optimized through weighted summation. For example, when the system predicts that the vehicle in front successfully changes lanes and avoids collision, 、 , the comprehensive reward drives the model to enhance the parameter weight of this scenario. This reward mechanism makes the prediction accuracy in complex scenarios improve by 5-8% per month (as shown in Equation Two), breaking through the limitations of traditional manual parameter adjustment: Dynamic reward function: (Equation Two); Where: ; ; ; is the weight coefficient, dynamically adjusted according to the scene (such as the urban scene , , ), highlighting the priority of safety or efficiency; is the state space (such as the current traffic scene, sensor data); is the action space (such as adjusting model parameters); .

[0060] From the algorithm layer, the problem of "poor adaptability to dynamic environment" is solved, and through the three-dimensional reward mechanism, the model autonomously learns the sudden behavior of traffic participants (such as pedestrians stopping suddenly), filling the extreme scenes not covered by traditional databases, and further shortening the prediction delay.

[0061] 2. Optimization algorithm: Existing optimization algorithms often focus on single target optimization, which cannot balance the contradiction between perception accuracy, matching accuracy, and computational resource consumption, and lack the ability to adapt to complex dynamic environments. This method combines double population particle swarm optimization (DPSO) with differential evolution algorithm (DE) to form a hybrid intelligent optimization algorithm, and introduces chaos mapping and Cauchy mutation strategy to build a new optimization framework.

[0062] The design of chaos mapping initializes the particle swarm, which utilizes the ergodicity and randomness of chaos system to break the dilemma of traditional particle swarm optimization algorithm that is easy to fall into local optimum by random initialization. By generating initial particle positions through Logistic chaos mapping, particles are more evenly distributed in the solution space, increasing the possibility of global search. For example, when optimizing physical field perception parameters, the particle swarm after chaos initialization can quickly explore the optimal combination of magnetic resistance sensor sampling frequency and acoustic sensor emission period, and the convergence speed is improved compared with traditional random initialization.

[0063] The Cauchy mutation strategy enhances the local search ability of the algorithm, which is different from the fixed step and direction of traditional mutation strategies. In the later stage of the algorithm, when the particle swarm gradually gathers around the local optimal solution, Cauchy mutation produces long-distance jumps with a high probability, avoiding premature convergence of the algorithm. When optimizing biological behavior pattern matching parameters, this strategy can help the algorithm jump out of the local optimal parameter combination and find a better behavior feature weight distribution scheme, improving the behavior pattern matching accuracy.

[0064] The construction of the multi-objective optimization function is based on a redefinition of the system's core requirements, simultaneously considering three key objectives: physical field perception accuracy, behavioral pattern matching accuracy, and computational resource consumption. Through Pareto frontier analysis, the system is able to find the optimal compromise solution for multiple objectives, rather than the optimal solution for a single objective. For example, when computing resources are limited, the system can automatically adjust parameters to reduce computing resource consumption while maintaining a certain level of perception accuracy and matching accuracy. This achieves efficient resource utilization and addresses the problem that traditional algorithms cannot balance multi-objective optimization.

[0065] Different from traditional single-objective optimization, this method establishes a three-dimensional objective function vector (see Formula 3). Based on the Pareto optimality theory, by minimizing the loss of perceptual accuracy , matching accuracy loss and computing resource consumption For example, when computing resources are tight, the function automatically adjusts the parameters ,exist 、 Under the constraints of Reduce by 25% (as shown in Formula 3), breaking through the resource bottleneck of traditional algorithms: Multi-objective optimization function vector representation: (Formula 3); in: , perceptual accuracy loss function; , matching accuracy loss function; , computing resource consumption function; , is the optimization parameter vector (such as sensor sampling frequency, model weight, etc.); Pareto optimality conditions: Make and ; in is the feasible solution space, is the current solution, For any other solution, is the objective function index ( Physical field perception accuracy, Behavior matching accuracy, Computing resource consumption); the benchmark value is the industry-recognized performance standard or historical best value.

[0066] Quantization solves the problem of "conflict between computing resources and prediction accuracy", and through Pareto frontier analysis, it realizes the reduction of resource consumption while maintaining accuracy, breaking the traditional algorithm "high calculation = high delay" cycle.

[0067] Online incremental learning is realized by using Flink real-time computing framework, which changes the defect that traditional batch learning mode cannot adapt to environmental changes in time. Flink can receive new data in real time in the form of stream processing. When road construction leads to temporary changes in traffic rules, or large-scale activities cause changes in crowd behavior patterns, the system can complete model parameter update within minutes, so that the prediction model quickly adapts to the new environment. Compared with the traditional batch learning mode, the response speed is improved, which greatly enhances the adaptability and stability of the system in dynamic complex environment.

[0068] Summary: Abandoning the traditional single function sensor combination, composite sensor unit is adopted, integrating magnetoresistance sensor array and acoustic resonance sensor; vehicle end design paper structure flexible circuit board composite antenna, using frequency spectrum modulation and OFDM technology to process signal. At the same time, construct the earth vehicle cooperative collection network, combined with unmanned aerial vehicle and vehicle camera to collect data. It solves the problems of high data collection delay, multi-source data fusion difficulty and data collection perspective limitation of traditional sensor. Magnetoresistance sensor high precision and three axis detection, acoustic resonance sensor phased array technology and adaptive filter, improve detection precision and range, reduce environmental interference; Composite antenna realizes efficient signal processing, detection distance is improved in complex environment; Earth vehicle cooperative collection obtains more comprehensive data, fills the gap of behavior motivation data.

[0069] Adopting distributed heterogeneous database architecture, combining block chain and graph database; Introducing reinforcement learning optimization module to dynamically adjust the database index structure; When building biological behavior pattern database, combine various technologies to analyze behavior data and establish emotion and behavior pattern mapping. It solves the problems of low data security, complex relationship processing, low query efficiency and lack of behavior motivation analysis in traditional database and behavior intention prediction. Blockchain ensures data tamper-proof, graph database efficiently processes complex relationship query, and complex query response time is shortened to milliseconds; Emotion and behavior pattern mapping makes the prediction model closer to the real scene, and the prediction ability of special scene is significantly improved.

[0070] A three-level probabilistic inference network is constructed, combined with multiple algorithms to process physical field and biological behavior data; the physical field data is analyzed by combining variational mode decomposition and Hilbert-Huang transform, and the biological behavior mode is processed by cloud-Markov chain model; the vehicle-mounted computing platform adopts a heterogeneous multi-core architecture and a task dynamic scheduling algorithm. The problems of large amount of calculation, difficulty in processing data fuzziness and uncertainty in traditional decision-making algorithm, and insufficient utilization of resources in vehicle-mounted computing platform are solved. The three-level network realizes multi-dimensional information fusion decision, predicts the motion intention in advance and outputs the probability distribution; the physical field and biological behavior data processing method deeply excavates the data characteristics; the heterogeneous multi-core architecture improves the utilization rate of computing resources, reduces energy consumption, and ensures the real-time and accuracy of prediction in complex scenarios.

[0071] An edge-cloud collaborative computing architecture is constructed, a double closed-loop control strategy is adopted to adjust the sensor and model parameters; combined with multiple optimization algorithms, a new optimization framework and function are designed; Flink real-time computing framework is adopted to realize incremental learning. The problems of high delay, untimely response of traditional feedback system, inability of optimization algorithm to balance multiple objectives and adapt to dynamic environment, and inability of learning mode to adapt to environmental changes in time are solved. Edge-cloud collaboration reduces data transmission delay, double closed-loop control improves sensor accuracy and model accuracy; mixed optimization algorithm balances multiple objectives, realizes efficient use of resources; Flink framework makes the model update every minute, and the adaptability and stability of the system are greatly enhanced.

Claims

1. A motion intention prediction method for dynamic fuzzy trajectories, characterized by: It includes sensor data acquisition, multi-source data fusion and motion intention decision-making, specifically: Build a physical field sensing network, using a composite sensor unit integrating a magnetoresistive sensor array and an acoustic resonance sensor, combined with spectrum modulation of the vehicle-side composite antenna, to achieve low-latency acquisition and transmission of physical field signals; Build a biological behavior pattern database, acquire traffic participant behavior data through a collaborative data collection network between ground and vehicle, and introduce micro-expression recognition technology to establish a probabilistic mapping model between emotions and behavior patterns; Construct a three-level probabilistic reasoning network to make decisions based on the integration of physical field characteristics, behavioral patterns, and environmental information through convolutional autoencoders, Bayesian networks, and DS evidence theory; An edge-cloud collaborative feedback mechanism is established to dynamically optimize sensor parameters and model parameters through adaptive PID algorithm and deep Q network.

2. The method for predicting motion intention of a dynamic fuzzy trajectory according to claim 1, wherein: The composite sensor unit integrates a magnetoresistive sensor and an acoustic resonance sensor into a buryable modular device, wherein the magnetoresistive sensor adopts a three-axis magnetoresistive chip and is equipped with a temperature compensation circuit.

3. The motion intention prediction method of a dynamic fuzzy trajectory according to claim 1, wherein: The vehicle-side composite antenna uses an origami-structured flexible circuit board and integrates a dual-channel spectrum modulator. It independently processes electromagnetic signals and acoustic signals through orthogonal frequency division multiplexing, and combines a pulse code modulation generator to enhance the target echo signal.

4. The method for predicting motion intention of a dynamic fuzzy trajectory according to claim 1, wherein: The collaborative data collection network between the ground and the vehicle includes a swarm of drones, vehicle-mounted cameras, and distributed collaborative positioning technology. The drones collect 120° wide-angle videos at an altitude of 100 meters and a speed of 15km / h. The vehicle-mounted cameras are equipped with dynamic aperture and optical image stabilization systems, and combine the spatiotemporal attention mechanism to extract behavioral features.

5. The method for predicting motion intention of a dynamic fuzzy trajectory according to claim 1, wherein: The three-level probabilistic reasoning network includes: The physical field feature extraction layer uses a convolutional autoencoder to reduce the dimension of the magnetic resistance and acoustic wave signals and remove noise; The behavior pattern matching layer calculates the probability distribution of movement intention through the Bayesian network; The dynamic decision-making layer corrects the intention probability by fusing environmental information based on DS evidence theory.

6. The method for predicting motion intention of a dynamic fuzzy trajectory according to claim 1, wherein: The biological behavior pattern database adopts a distributed heterogeneous architecture. The blockchain layer ensures that the data cannot be tampered with through a practical Byzantine fault-tolerant algorithm. The graph database layer constructs a relationship network including behavior nodes, environmental factor nodes, and behavior motivation nodes, and combines the reinforcement learning optimization module to dynamically adjust the index.

7. The method for predicting motion intention of a dynamic fuzzy trajectory according to claim 1, wherein: The edge-cloud collaborative feedback mechanism adopts dual closed-loop control: The inner loop dynamically adjusts the sensor parameters through the adaptive PID algorithm, and jointly controls the PID parameters based on the error gradient, cumulative amount and acceleration; The outer loop optimizes model parameters through a deep Q network combined with a dynamic reward function, with safety, accuracy, and efficiency as the three-dimensional reward dimensions.

8. The method for predicting motion intention of a dynamic fuzzy trajectory according to claim 1, wherein: The feedback optimization mechanism introduces a hybrid intelligent algorithm of dual-population particle swarm optimization and differential evolution algorithm, combined with chaotic map initialization and Cauchy mutation strategy, to balance the physical field perception accuracy, behavior matching accuracy and computing resource consumption through a multi-objective optimization function.

Images