Automobile intelligent image processing system and method based on perception algorithm model

Abstract

The application provides an intelligent image processing system and method for a car based on a perception algorithm model, and the method comprises the following steps: S1. spatio-temporal reference double anchoring and dynamic intrinsic extrinsic parameter calibration of a vehicle-mounted image acquisition node; S2. multi-node image heterogeneous domain normalization and adaptive preprocessing based on a self-adaptive kernel regression non-local mean denoising algorithm; S3. hierarchical image feature extraction and semantic anchoring based on a graph neural network dynamic feature interaction network; S4. cross-node and cross-frame feature mutual checking and pseudo-feature elimination; S5. full-scene semantic completion and dynamic target trajectory prediction based on a variational autoencoder trajectory prediction model; S6. dynamic lightweight adaptation and algorithm power adaptive scheduling of the perception algorithm model; and S7. risk scene grading identification and image targeted enhancement output based on semantics and trajectories. The application provides stable, accurate and efficient vehicle-mounted image perception support for intelligent driving of a car, and improves the safety and adaptability of environmental perception of intelligent driving.

Description

Technical Field

[0001] This invention relates to the field of automotive intelligent driving environment perception technology, and in particular to an automotive intelligent image processing system and method based on perception algorithm models. Background Technology

[0002] With the rapid popularization of intelligent driving technology in automobiles, in-vehicle image perception systems have become the core data source for vehicle environment recognition, risk warning, and autonomous driving decision-making. Most existing in-vehicle image processing solutions adopt a segmented pipeline architecture, sequentially completing image acquisition, basic denoising, feature extraction, target detection, and trajectory prediction. Such conventional solutions have many inherent technical defects in real-world complex in-vehicle scenarios.

[0003] In existing technologies, the image preprocessing stage and the perception algorithm stage are independent of each other. Denoising, normalization, and other processes are performed only based on the pixel features of the image itself, and cannot be adaptively adjusted by combining subsequent perception feature weights. This easily leads to problems such as excessive blurring of key features or incomplete noise suppression. The feature extraction stage mostly uses conventional convolutional neural networks, which can only extract grid-like pixel features according to fixed rules. It is difficult to model irregular targets and the spatial and semantic relationships of cross-camera related features, resulting in poor targeting and robustness of feature extraction. The trajectory prediction stage mostly uses deterministic linear models, which are calculated only based on historical trajectory data and do not integrate perception features and environmental semantic information. The prediction results are singular and unreliable.

[0004] Meanwhile, existing solutions generally suffer from one-way data transmission between algorithms and lack of bidirectional interactive collaboration. Preprocessing results cannot be dynamically optimized based on perception feedback, perception features cannot be adaptively adjusted in conjunction with trajectory prediction confidence, and trajectory prediction is difficult to improve accuracy by relying on perception features. In addition, existing solutions lack a full-link closed-loop calibration mechanism. The spatiotemporal reference of the vehicle camera is prone to drift due to vehicle vibration and temperature changes. False features cannot continuously provide feedback to optimize preprocessing and perception models, and computing resources cannot be dynamically adapted to scene complexity. In highly complex and computationally demanding scenarios, perception delays and accuracy degradation are likely to occur, making it difficult to meet the stringent requirements of intelligent driving for image processing accuracy, real-time performance, and scene adaptability. Summary of the Invention

[0005] This invention provides an intelligent automotive image processing system and method based on a perception algorithm model. It aims to achieve synergistic optimization of image preprocessing and perception feature extraction, and mutual enhancement of feature extraction and trajectory prediction through deep fusion and bidirectional interaction of algorithms. This improves the feature preservation accuracy, target detection reliability, and trajectory prediction accuracy of in-vehicle image processing. Simultaneously, through dynamic spatiotemporal benchmark calibration, adaptive computing power scheduling, and end-to-end closed-loop self-optimization, it addresses issues such as image perception drift, insufficient computing power, and poor robustness in complex in-vehicle scenarios. Ultimately, it provides stable, accurate, and efficient in-vehicle image perception support for intelligent driving, enhancing the safety and adaptability of intelligent driving environment perception.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The automotive intelligent image processing method based on perception algorithm models includes: S1. Establish a globally unified time reference and vehicle body reference coordinate system, complete the dynamic calibration of intrinsic and extrinsic parameters corresponding to image acquisition, and generate a unified spatiotemporal reference parameter package for the entire link; S2. Based on the spatiotemporal reference parameter package, the vehicle-mounted raw image data is sequentially processed for distortion correction, cross-node field of view division, heterogeneous domain normalization and adaptive denoising, generating a preprocessed output data package; S3. Based on the spatiotemporal reference parameter package and the preprocessed output data package, image hierarchical feature extraction is completed through the graph neural network dynamic feature interaction network. Static semantic segmentation and dynamic target detection are completed based on the extracted features, and feature extraction output data package is generated. S4. Based on the spatiotemporal reference parameter package, preprocessed output data package and feature extraction output data package, perform double cross-verification on static features and dynamic features respectively, eliminate false features, and generate valid feature data packages after verification. S5. Based on the spatiotemporal reference parameter package, feature extraction output data package and effective feature data package, complete the spatiotemporal continuous completion of static semantics and multi-target tracking of dynamic targets. Through the variational autoencoder trajectory prediction model, complete the multi-temporal trajectory prediction of dynamic targets and generate semantic and trajectory output data packages. S6. Based on semantic and trajectory output data packets, combined with real-time computing power usage data and vehicle real-time driving speed, complete the scene complexity and computing power load status assessment, and generate dynamic adaptation strategy parameters; S7. Based on semantic and trajectory output data packets and dynamic adaptation strategy parameters, combined with real-time vehicle body status data, complete the risk scene classification identification, and perform differentiated image enhancement and output based on the risk level.

[0007] This specification also includes: S8. Based on external truth reference data and combined with the output results of each step in the entire chain, complete the closed-loop verification of the entire chain processing results, generate corresponding calibration and update parameters, and complete the entire chain self-optimization; the closed-loop verification includes three types: spatiotemporal reference parameter closed-loop verification, preprocessing parameter closed-loop optimization, and perception algorithm model incremental self-calibration; the spatiotemporal reference parameter closed-loop verification calculates the reprojection error of static feature points based on truth reference data, and generates a calibration trigger signal to be fed back to S1 when it exceeds a preset threshold; the preprocessing parameter closed-loop optimization generates parameter update signals to be fed back to S2 based on pseudo-feature statistical information; the perception algorithm model incremental self-calibration completes feature error analysis based on truth reference data, constructs an incremental training sample set to complete model fine-tuning, and the generated model weight update parameters are fed back to S3.

[0008] In this specification, in S1, the vehicle body reference coordinate system is a three-dimensional rectangular coordinate system established with the rear axle center of the vehicle as the origin, the vehicle's driving direction as the positive X-axis, the vehicle's vertical upward direction as the positive Z-axis, and the vehicle's left lateral direction as the positive Y-axis. The global unified time reference is established by accurately marking the global unified timestamp onto the frame header of each frame of the original vehicle image data to ensure that the time synchronization error of all image data does not exceed 1 millisecond. The dynamic calibration of intrinsic parameters is completed based on the static feature constraints of the natural scene and the bundle adjustment method, while the dynamic calibration of extrinsic parameters is completed based on the vehicle pose recursion and cross-node feature matching.

[0009] In this specification, in step S2, the adaptive denoising is performed using an adaptive kernel regression nonlocal mean denoising algorithm. This algorithm simultaneously integrates the spatial distance constraints of local pixels and the gray-level similarity constraints of nonlocal pixels to calculate the denoising weights pixel by pixel, thereby completing image denoising. The denoising weights are adaptively adjusted based on the graph node feature weight matrix generated in step S3. The denoising intensity is reduced in areas with high feature weights to preserve detailed features, while the denoising intensity is increased in areas with low feature weights to suppress noise interference.

[0010] In this specification, in S3, the execution process of the graph neural network dynamic feature interaction network is as follows: superpixels are generated from the preprocessed image data; an initial dynamic graph structure is constructed with each superpixel as a graph node; the initial feature vector of each graph node is extracted; the initial graph edge weights are calculated; and multiple rounds of graph node feature iteration updates are completed through a graph attention mechanism to generate the final graph node feature vector set and graph node feature weight matrix; the initial graph edge weights are dynamically adjusted based on the trajectory prediction confidence matrix generated in S5 to improve the feature interaction intensity of graph nodes corresponding to high-confidence dynamic targets.

[0011] In this specification, in S4, the dual mutual verification performed on static features is cross-node spatial mutual verification and cross-frame temporal mutual verification, which determines the validity of static features based on feature similarity, semantic category consistency, and inter-frame feature trajectory stability; the dual mutual verification performed on dynamic features is cross-node spatial mutual verification and motion constraint consistency verification, which determines the validity of dynamic features based on spatial position consistency, target category consistency, and vehicle scene kinematic constraints; the pseudo-feature region mask generated during the verification process is fed back to S2 for targeted region enhancement processing in subsequent image frames.

[0012] In this specification, in S5, the execution process of the variational autoencoder trajectory prediction model is as follows: the historical motion trajectory information of the dynamic target, the surrounding static semantic environment information, and the corresponding graph node feature information are fused to construct the encoder input vector. The encoder maps the input vector to the latent space, and the latent variables are sampled using the reparameterization technique. Then, the decoder outputs the predicted trajectory points of the dynamic target in the future multi-time series, and at the same time generates the trajectory prediction confidence matrix. The trajectory prediction confidence matrix is ​​fed back to S3 to dynamically adjust the graph edge weights of the dynamic feature interaction network of the graph neural network.

[0013] In this specification, in S6, the scene complexity score is calculated by weighting the dynamic target quantity weight, scene semantic category richness weight, vehicle speed weight, and image imaging quality weight; the computing power load state is divided into three categories: high computing power load state, medium computing power load state, and low computing power load state; according to the scene complexity score and computing power load state, corresponding dynamic adaptation strategy parameters are generated, and the dynamic adaptation strategy parameters are fed back to S3 to adjust the graph neural network structure, feature extraction rules, and inference accuracy of subsequent image frames.

[0014] In this manual, S7 predefines four risk levels: safety level, alert level, warning level, and emergency level. Each risk level corresponds to a preset trigger condition. The risk level determination process is as follows: predict the vehicle's future trajectory based on real-time vehicle body status data, calculate the collision time and minimum safe distance by combining the predicted trajectory of dynamic targets, and identify risk factors by combining static semantic information to complete the risk level determination. The higher the risk level, the higher the image magnification and enhancement intensity of the corresponding risk area, and the higher the output priority. The identified risk area coordinate set is fed back to S2 for targeted area enhancement processing in subsequent image frames.

[0015] The automotive intelligent image processing system based on the perception algorithm model, applying the automotive intelligent image processing method based on the perception algorithm model described above, comprises: a spatiotemporal reference calibration module for performing operation S1 and generating a spatiotemporal reference parameter package; an adaptive preprocessing module for performing operation S2 and generating a preprocessing output data package; a graph neural network feature extraction module for performing operation S3 and generating a feature extraction output data package; a pseudo-feature verification and removal module for performing operation S4 and generating a verified valid feature data package; a semantic completion and trajectory prediction module for performing operation S5 and generating semantic and trajectory output data packages; a computing power adaptive scheduling module for performing operation S6 and generating dynamic adaptation strategy parameters; a risk classification and image output module for performing operation S7 and completing risk identification and image enhancement output; and a full-link closed-loop calibration module for performing operation S8 and completing full-link closed-loop verification and self-optimization.

[0016] In summary, the present invention has at least the following beneficial effects: Through the bidirectional interaction between the adaptive kernel regression nonlocal mean denoising algorithm and the graph neural network dynamic feature interaction network, the present invention enables image preprocessing to adaptively adjust the denoising intensity according to the perceptual feature weights, completely suppressing noise while fully preserving key edge and texture features, avoiding the feature loss problem caused by the disconnect between conventional preprocessing and perception; through the bidirectional collaboration between the graph neural network dynamic feature interaction network and the variational autoencoder trajectory prediction model, the trajectory prediction confidence can dynamically optimize the feature interaction weights, improving the feature extraction accuracy of key targets, while perceptual features can in turn enhance the multimodal accuracy of trajectory prediction, achieving mutual benefit between feature extraction and trajectory prediction.

[0017] This invention effectively eliminates false features caused by reflections, shadows, and noise through cross-node and cross-frame dual feature verification, improving the reliability of static semantic segmentation and dynamic target detection. By dynamically adapting to scene complexity and computational load, it can flexibly adjust model structure and computational consumption while ensuring core perception accuracy, adapting to the real-time requirements of different vehicle scenarios. Through a full-link closed-loop calibration mechanism, it can automatically correct camera spatiotemporal reference drift, continuously optimize preprocessing parameters and perception model weights, and improve the long-term stability and robustness of the solution. Based on risk level-based differentiated image enhancement, it can accurately highlight information in high-risk areas, improving the intuitiveness and effectiveness of risk warnings. Overall, it achieves a technological upgrade in vehicle image processing from passive processing to active adaptation, and from independent segmentation to full-link collaboration. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of an intelligent image processing method for automobiles based on a perception algorithm model.

[0019] Figure 2 This is a schematic diagram illustrating the interaction between the adaptive kernel regression nonlocal means denoising algorithm and the graph neural network dynamic feature interaction network.

[0020] Figure 3 This is a schematic diagram illustrating the interaction between a graph neural network dynamic feature interaction network and a variational autoencoder trajectory prediction model.

[0021] Figure 4 This is a schematic diagram illustrating the interaction between the variational autoencoder trajectory prediction model and the graph neural network dynamic feature interaction network. Detailed Implementation

[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] like Figure 1 As shown, this embodiment provides an intelligent automotive image processing method based on a perception algorithm model. With a core architecture of end-to-end bidirectional collaboration and closed-loop self-optimization, it sequentially executes: dual anchoring and dynamic calibration of the vehicle image spatiotemporal reference, multi-node image adaptive preprocessing, dynamic graph structure feature extraction, cross-node and cross-frame pseudo-feature removal, multi-modal trajectory prediction, dynamic computing power adaptation, risk-level image enhancement, and end-to-end closed-loop calibration. This method selects three types of algorithms as core support: adaptive kernel regression nonlocal mean denoising algorithm, graph neural network dynamic feature interaction network, and variational autoencoder trajectory prediction model. These three algorithms form a bidirectional interactive relationship between each other (see reference). Figure 2 , Figure 3 and Figure 4 The preprocessing denoising algorithm adaptively adjusts the feature weights of graph nodes received from the feature extraction stage, the feature extraction network optimizes the graph edge weights by receiving the confidence matrix from the trajectory prediction stage, and the trajectory prediction model improves prediction accuracy by relying on the graph node features extracted from the feature extraction stage. At the same time, through spatiotemporal benchmark dynamic calibration, pseudo-feature feedback verification, adaptive computing power scheduling and full-link incremental calibration, a complete closed loop of "processing-verification-feedback-optimization" is formed, ultimately achieving high-quality processing, accurate perception and efficient risk identification of vehicle-mounted multi-source images.

[0024] S1. Spatiotemporal reference dual anchoring and dynamic intrinsic extrinsic parameter calibration of vehicle-mounted image acquisition nodes After the vehicle is powered on, this step starts simultaneously. First, it obtains the real-time pose data of the vehicle body and the globally unified timestamp from the spatiotemporal reference unit. Then, it obtains the factory calibration intrinsic parameters and initial extrinsic parameters of each vehicle image acquisition node from the preset storage area of ​​the image processing domain controller. Finally, it obtains the initial continuous frame image data from each vehicle image acquisition node. At the same time, it reserves a callback interface to receive the calibration trigger signal fed back by S8, and completes the establishment and dynamic calibration of the unique spatiotemporal reference of the entire link, providing a unified time and space reference for all subsequent links.

[0025] The hardware devices involved in this invention are all existing technologies, such as the vehicle image acquisition node using the Sony IMX390 image sensor module, the spatiotemporal reference unit using the Novatek OEM7720 vehicle integrated navigation unit, and the image processing domain controller using the NVIDIA Drive Orin vehicle high-computing power domain controller.

[0026] S1.1 Establishment of a Global Hard Synchronization Time Base. For each vehicle-mounted image acquisition node, the globally unified timestamp output by the spatiotemporal reference unit is precisely marked onto the frame header of each frame of image data output by each vehicle-mounted image acquisition node through the hardware synchronization interface of the image processing domain controller. The timestamp marking accuracy is no less than 1 millisecond, ensuring that the time synchronization error of image data from all vehicle-mounted image acquisition nodes does not exceed 1 millisecond. After this step is completed, image frame marking rules with unified timestamps and a global time base are generated, completely solving the problem of asynchronous time acquisition of multi-node images and avoiding spatial positioning errors and feature matching failures caused by time deviations in high-speed driving scenarios.

[0027] S1.2 Initial Anchoring of the Global Spatial Coordinate System. This step uses the vehicle's body reference coordinate system as the unique spatial coordinate system for the entire link. This coordinate system establishes a three-dimensional Cartesian coordinate system with the vehicle's rear axle center as the origin, the vehicle's driving direction as the positive X-axis, the vehicle's vertical upward direction as the positive Z-axis, and the vehicle's left lateral direction as the positive Y-axis. This step converts the factory-set initial extrinsic parameters of each vehicle-mounted image acquisition node into an extrinsic parameter matrix relative to the vehicle's body reference coordinate system. The extrinsic parameter matrix contains rotation matrices and translation vectors, used to describe the transformation relationship between the optical center coordinate system of the vehicle-mounted image acquisition node and the vehicle's body reference coordinate system. After this step is completed, an initial extrinsic parameter matrix set is generated, providing a unified spatial reference for subsequent spatial alignment, feature projection, and cross-node feature matching throughout the entire link.

[0028] S1.3 Dynamic calibration of intrinsic parameters based on static feature constraints of natural scenes. For each vehicle-mounted image acquisition node, at least 30 frames of image data of the same scene are continuously acquired. The AKAZE feature detection algorithm is used to extract static feature points in each frame. Based on the matching trajectory of static feature points in multiple frames, the bundle adjustment method is used to optimize and solve the real-time intrinsic parameters of the vehicle-mounted image acquisition node. The intrinsic parameters include the intrinsic parameter matrix, radial distortion coefficient, and tangential distortion coefficient. The optimized real-time intrinsic parameters replace the factory-calibrated intrinsic parameters, and the changes in intrinsic parameters are recorded. If the change rate of the intrinsic parameter matrix elements is not less than 0.5% or the change rate of the distortion coefficient is not less than 2%, a distortion correction weight update signal is generated for parameter adjustment in the subsequent distortion correction stage of S2. This step solves the problem of intrinsic parameter drift caused by vehicle bumps, changes in ambient temperature, and lens aging, and avoids subsequent feature extraction errors caused by distortion correction failure.

[0029] S1.4 Dynamic calibration of extrinsic parameters based on vehicle pose recursion and cross-node feature matching. Based on the real-time vehicle pose data output by the spatiotemporal reference unit, the pose change of the vehicle between consecutive frames is recursively calculated. Simultaneously, AKAZE static feature points are extracted from the overlapping field-of-view image data of adjacent vehicle-mounted image acquisition nodes, and cross-node feature point matching is completed. Combining the vehicle pose recursion results and the cross-node feature point matching results, a reprojection error constraint equation is constructed, and the Levenberg-Marquardt algorithm is used to optimize and solve the real-time extrinsic parameter matrix of each vehicle-mounted image acquisition node relative to the vehicle body reference coordinate system. The optimized real-time extrinsic parameter matrix replaces the initial extrinsic parameter matrix, and the change in the extrinsic parameter matrix is ​​recorded. If the angular deviation of the rotation matrix is ​​not less than 0.2° or the deviation of the translation vector is not less than 5 mm, a spatial alignment weight update signal is generated for parameter adjustment in the subsequent cross-node field-of-view alignment step of S2. This step solves the problem of extrinsic parameter offset caused by vehicle installation deformation and vibration, ensuring the spatial coordinate consistency of multi-node images.

[0030] S1.5 Spatiotemporal Reference Parameter Package Generation and Callback Mechanism Establishment. This step integrates the outputs of the preceding steps to generate a unified spatiotemporal reference parameter package across the entire chain. This package is stored in the shared memory area of ​​the image processing domain controller for real-time access by all subsequent steps. The spatiotemporal reference parameter package includes a global time reference, real-time intrinsic parameter sets for each vehicle-mounted image acquisition node, real-time extrinsic parameter matrix sets, intrinsic parameter changes, and extrinsic parameter changes. Simultaneously, a parameter update callback mechanism is established. Upon receiving the calibration trigger signal from S8, the calibration process from S1.3 to S1.4 is immediately re-executed to update the spatiotemporal reference parameter package, achieving dynamic maintenance of the spatiotemporal reference throughout its entire lifecycle.

[0031] S2. Multi-node image heterogeneous domain normalization and adaptive preprocessing based on adaptive kernel regression nonlocal mean denoising algorithm This step, based on the spatiotemporal reference parameter package, distortion correction weight update signal, and spatial alignment weight update signal generated in S1, performs distortion correction, field of view division, heterogeneous domain normalization, adaptive denoising, and targeted enhancement processing on the raw image data with unified timestamps output by each vehicle-mounted image acquisition node. Simultaneously, it receives the graph node feature weight matrix fed back from S3, the risk region coordinate set fed back from S7, and the pseudo-feature region mask and preprocessing parameter update signal fed back from S8, realizing dynamic optimization of the preprocessing stage and providing high-quality, highly consistent image data for the subsequent feature extraction stage.

[0032] S2.1 Regional Distortion Correction Based on Real-Time Intrinsic Parameter Set. For each frame of raw image data, a pixel-level distortion mapping table is constructed based on the real-time intrinsic parameter set output by S1. Simultaneously, the interpolation weights for distortion correction are adaptively adjusted according to the changes in intrinsic parameters and the distortion correction weights. If the changes in intrinsic parameters do not exceed a preset threshold, bilinear interpolation is used to complete distortion correction; if the changes exceed the preset threshold, bicubic interpolation is used to improve the accuracy of distortion correction. After distortion correction is completed, distortion-free image data is generated, and the initial projection coordinates of each pixel in the vehicle's reference coordinate system are recorded, providing basic data for subsequent full-link spatial matching and feature anchoring.

[0033] S2.2 Cross-node field-of-view region partitioning based on real-time extrinsic parameter matrix set. For distortion-free image data of all vehicle-mounted image acquisition nodes, based on the real-time extrinsic parameter matrix set output from S1, the spatial coverage area of ​​the field of view of each vehicle-mounted image acquisition node in the vehicle's reference coordinate system is calculated. All field of view ranges are divided into three types of regions, and each pixel is labeled with its region type, generating region mask labels. Specifically, these include non-overlapping field of view regions, two-node overlapping field of view regions, and multi-node overlapping field of view regions. Non-overlapping field of view regions refer to those covered by only a single vehicle-mounted image acquisition node, two-node overlapping field of view regions refer to those covered by two vehicle-mounted image acquisition nodes simultaneously, and multi-node overlapping field of view regions refer to those covered by three or more vehicle-mounted image acquisition nodes simultaneously. The region partitioning results in this step will provide a basis for subsequent differentiated normalization processing and feature extraction weight allocation.

[0034] S2.3 Heterogeneous Domain Normalization Processing Based on Region Type and Imaging Quality Evaluation. For each frame of distortion-free image data, the imaging quality evaluation indicators of the entire image are first calculated, including signal-to-noise ratio, exposure, color temperature difference, and blur. Then, combined with region mask labels, differentiated normalization processing is performed on different types of regions to ensure imaging consistency across nodes. For non-overlapping field-of-view regions, adaptive gamma correction and automatic white balance processing are adopted. Based on the gray-level distribution within the region, the gamma value and white balance gain are adaptively adjusted to normalize the image data within the region to the preset standard dynamic range and color temperature range. For dual-node overlapping field-of-view regions, the image data of the corresponding regions of the two vehicle-mounted image acquisition nodes are first subjected to joint histogram matching of brightness and chromaticity to normalize the image data of the two regions to the same brightness and chromaticity distribution. Then, bilateral filtering is performed to remove noise while preserving edge features. For multi-node overlapping field-of-view regions, the average brightness and average chromaticity of the image data of the corresponding regions of all nodes are first used as a benchmark to perform node-by-node brightness-chromaticity linear mapping. Then, guided filtering is used to remove noise while preserving the consistency of edge features. After processing, normalized image data and regional imaging quality evaluation parameters are generated, providing support for subsequent denoising and feature extraction.

[0035] S2.4 Adaptive Denoising Based on Adaptive Kernel Regression Nonlocal Means Denoising Algorithm. This step employs an adaptive kernel regression nonlocal means denoising algorithm to denoise the normalized image data. This algorithm differs from conventional nonlocal means denoising algorithms by integrating spatial distance constraints of local pixels with grayscale similarity constraints of nonlocal pixels. It can also receive feature weights from subsequent S3 feedback for adaptive adjustment. This solves the pain points of conventional denoising algorithms in complex vehicle scenarios, which either suffer from severe noise residue or excessive blurring of key features. It is the core carrier for the collaborative optimization of preprocessing and perception stages in this solution.

[0036] The core computational process of the algorithm is as follows: ; in, q represents the grayscale value of pixel p after denoising, where p represents the coordinates of the current pixel to be processed, and q represents the coordinates of any pixel within the search window. This represents a search window centered on pixel p, with a window size of 21×21. This represents the gray value of pixel q in the normalized image. This represents the adaptive weight of pixel q for pixel p.

[0037] The formula for calculating adaptive weights is as follows: ; in, This represents a normalization constant used to ensure that the sum of the weights of all associated pixels is 1. This represents the weighted Euclidean distance between two neighboring windows. This represents a neighborhood window centered at pixel p, with a window size of 7×7. This represents a neighborhood window centered at pixel q. The Gaussian kernel parameter represents the similarity of gray levels. A smoothing parameter representing grayscale similarity. This represents the spatial Euclidean distance between two pixels. Gaussian kernel parameters representing spatial distance, A smoothing parameter representing spatial distance.

[0038] The formula for calculating the normalization constant is as follows: ; To achieve an adaptive balance between denoising intensity and feature preservation, this algorithm introduces the graph node feature weight matrix fed back by S3, and adjusts the smoothing parameters for gray-level similarity. Dynamic adjustments are made, using the following formula: ; in, This represents the initial value of the grayscale similarity smoothing parameter, which is set to 15. This represents the adjustment factor, set to 0.3. This represents the average value of the graph node feature weights corresponding to pixel p, which is output by the dynamic feature interaction network of the S3 graph neural network. When the graph node feature weight is high, it indicates that the region is a key feature region, and the smoothing parameter is reduced accordingly, the denoising intensity is weakened, and the detail features are preserved to the greatest extent. When the graph node feature weight is low, it indicates that the region is a non-key background region, and the smoothing parameter is increased accordingly, the denoising intensity is enhanced, and noise interference is effectively suppressed.

[0039] The model training process of this algorithm is as follows: Training is performed using the publicly available vehicle scene dataset BDD100K, which contains 100,000 labeled vehicle images covering different weather conditions, lighting conditions, and road scenarios. The training process employs unsupervised learning, optimizing the peak signal-to-noise ratio (PSNR) between the original and denoised images, and using the Adam optimizer to iteratively update the Gaussian kernel parameters. , With initial smoothing parameters , The training batch size is set to 16, the learning rate is set to 0.001, and the number of iterations is set to 500,000. After training, the optimal parameters are stored in the preset storage area of ​​the image processing domain controller for the algorithm to call in real time.

[0040] During the algorithm application process, the trained optimal parameters are first loaded from the preset storage area. Then, for each frame of normalized image data, adaptive weights are calculated pixel by pixel to complete the denoising process. At the same time, the graph node feature weight matrix fed back by S3 is received in real time, and the smoothing parameters of gray-level similarity are updated according to the dynamic adjustment formula to realize the region-by-region adaptive control of the denoising intensity, and finally the denoised image data is generated.

[0041] S2.5 is a region-specific enhancement process based on multi-stage feedback signals. This step receives the risk area coordinate set output from S7, the pseudo-feature region mask output from S8, and the preprocessing parameter update signal. For the image regions corresponding to these coordinates, targeted enhancement processing is performed, including contrast enhancement, edge sharpening, and noise suppression. The enhanced region data replaces the corresponding regions in the original denoised image data, generating the final preprocessed image data. Simultaneously, the enhancement processing parameters are fed back to the normalization processing stage in S2.3 to update the processing weights for similar regions, achieving continuous optimization of the preprocessing strategy.

[0042] S2.6 Preprocessing Output Data Package Generation. This step integrates all the output results from the previous steps to generate a preprocessing output data package, which is stored in the shared memory area of ​​the image processing domain controller for real-time access in subsequent feature extraction steps. The preprocessing output data package includes preprocessed image data with a unified timestamp, region mask labels, the vehicle body reference coordinate system projection coordinates corresponding to each pixel, region imaging quality evaluation parameters, enhancement processing parameters, and intermediate weight parameters of the denoising algorithm.

[0043] S3. Hierarchical Image Feature Extraction and Semantic Anchoring Based on Graph Neural Network Dynamic Feature Interaction Network This step, based on the preprocessed image data, region mask labels, and pixel-level spatial projection coordinates generated in S2, combined with the spatiotemporal reference parameter package output by S1, completes hierarchical image feature extraction, dynamic graph structure construction, feature interaction updates, spatiotemporal anchoring, and semantic target detection. Simultaneously, it receives the trajectory prediction confidence matrix from S5, the dynamic adaptation strategy parameters from S6, and the perception algorithm model update weight parameters from S8, achieving dynamic optimization and closed-loop calibration in the feature extraction stage. The generated graph node feature weight matrix is ​​also fed back to the denoising algorithm in S2, forming a bidirectional collaborative optimization between the preprocessing and perception stages. The graph neural network dynamic feature interaction network used in this step differs from the grid structure feature extraction mode of conventional convolutional neural networks. It constructs a dynamic graph structure using superpixels as graph nodes, which can flexibly model the spatial, semantic, and motion relationships between features. Furthermore, it can dynamically adjust the feature interaction weights through trajectory prediction feedback, solving the problem of insufficient feature extraction capabilities of conventional convolutional networks for irregular vehicle targets and cross-node correlations. It is the key hub for the perception core and multi-algorithm collaboration of this solution.

[0044] S3.1 Branch Predefinition and Weight Adaptive Initialization of the Perception Algorithm Model. This step uses a graph neural network dynamic feature interaction network as the backbone architecture, predefining three interrelated feature processing branches: static environment semantic branch, dynamic target feature branch, and detail texture feature branch. These three branches share the graph neural network backbone structure and are pre-trained based on an in-vehicle scene dataset, initializing the weight parameters of each branch. Simultaneously, based on the region mask labels and region imaging quality evaluation parameters output from S2, differentiated feature processing weights are set for the three branches. For non-overlapping field-of-view regions, the weight of the dynamic target feature branch is increased to prioritize the feature extraction accuracy of dynamic targets; for overlapping field-of-view regions, the weights of the static environment semantic branch and the detail texture feature branch are increased to prioritize the accuracy of spatial alignment and feature matching. If model update weight parameters are received from S8, the initial weights of the corresponding branches are directly replaced; if dynamic adaptation strategy parameters are received from S6, the number of superpixels, iterations, and branch channels of the graph structure are adjusted according to the strategy to achieve a dynamic balance between model accuracy and computational power consumption.

[0045] S3.2 Superpixel Generation and Initial Graph Structure Construction. The preprocessed image data output from S2 is used to generate superpixels using a simple linear iterative clustering algorithm. The initial number of superpixels is set to 500, with each superpixel corresponding to an independent graph node. The spatial distance between superpixels is calculated, and adjacent superpixels with a spatial distance of less than 30 pixels are connected to construct the initial graph edge set, generating the initial graph structure. ,in This represents the initial set of graph nodes. This represents the initial edge set of the graph.

[0046] S3.3 Initial Graph Node Feature Vector Extraction. For each superpixel, the appearance feature vector, spatial feature vector, and initial semantic feature vector are extracted and concatenated to generate the initial graph node feature vector, providing basic data for subsequent feature interactions. The appearance feature vector is composed of the mean, variance, maximum, and minimum values ​​of the grayscale, color, and gradient values ​​of all pixels within the superpixel, with a dimension of 64. The spatial feature vector is composed of the coordinates of the superpixel center in the vehicle's reference coordinate system, the area, perimeter, and shape factor of the superpixel, with a dimension of 32. The initial semantic feature vector is composed of shallow features extracted by the lightweight convolutional neural network MobileNetV3, with a dimension of 128. The calculation formula for the initial graph node feature vector is as follows: ; in, This represents the feature vector of the initial graph node corresponding to the i-th superpixel. This represents a vector concatenation operation. This represents the appearance feature vector of the i-th superpixel. This represents the spatial feature vector of the i-th superpixel. This represents the initial semantic feature vector of the i-th superpixel.

[0047] S3.4 Initial Graph Edge Weight Calculation and Dynamic Update Mechanism. First, the weights of the initial graph edges are calculated. These weights are a weighted sum of three components: appearance similarity, spatial proximity, and initial semantic relevance. The calculation formula is as follows: ; in, This represents the graph edge weight between the i-th and j-th superpixels in the initial state. The weighting coefficient representing appearance similarity is initially set to 0.3. The weighting coefficient representing spatial proximity is initially set to 0.3. The weighting coefficient representing semantic relevance is initially set to 0.4. This represents the appearance similarity between the i-th and j-th superpixels, with a value ranging from 0 to 1. It is calculated from the cosine similarity of the appearance feature vectors of the two superpixels. This represents the spatial proximity between the i-th and j-th superpixels, with values ​​ranging from 0 to 1, and is calculated using a Gaussian function of the spatial distance between the centers of the two superpixels. This represents the initial semantic correlation between the i-th superpixel and the j-th superpixel, with a value ranging from 0 to 1, and is calculated from the cosine similarity of the initial semantic feature vectors of the two superpixels.

[0048] To achieve dynamic optimization of feature interactions, this algorithm introduces the trajectory prediction confidence matrix fed back by S5, and dynamically updates the graph edge weights in each iteration. The update formula is as follows: ; in, Let represent the graph edge weight between the i-th and j-th superpixels at the t-th iteration, where t represents the iteration number, and the total number of iterations is set to 3. This represents the adjustment factor, set to 0.5. This represents the trajectory prediction confidence level corresponding to the i-th superpixel. This represents the trajectory prediction confidence level corresponding to the j-th superpixel. Both confidence level values ​​are output by the variational autoencoder trajectory prediction model of S5. When the trajectory prediction confidence level corresponding to a superpixel is high, it indicates that the region contains a high-risk dynamic target. The corresponding graph edge weights are increased accordingly, and the strength of feature interaction is enhanced, which can effectively improve the feature extraction accuracy of key targets and achieve bidirectional collaboration between trajectory prediction and feature extraction.

[0049] S3.5 Iterative Update of Graph Node Features Based on Graph Attention Mechanism. This step employs a graph attention network mechanism to complete multiple rounds of iterative updates of graph node features. Before each iteration, the graph edge weights are dynamically updated first, and then the feature update calculation is performed. The update formula is as follows: ; in, Let represent the graph node feature vector of the i-th superpixel after the t-th iteration. Represents the ReLU activation function. Let represent the set of neighbor nodes of the i-th superpixel at the t-th iteration. The set of neighbor nodes consists of adjacent nodes with edge weights greater than 0.2. Let represent the attention weight of the j-th superpixel to the i-th superpixel at the t-th iteration. Let represent the learnable weight matrix at the t-th iteration, with dimensions set to 256×256.

[0050] The formula for calculating attention weights is as follows: ; in, This represents the learnable parameter vector of the graph attention mechanism, with a dimension of 256. This represents a leaky ReLU activation function with a leak coefficient set to 0.2.

[0051] After three rounds of iteration, the final graph node feature vector set is generated. Simultaneously, a graph node feature weight matrix is ​​generated. The graph node feature weights are calculated from the 2-norm of the graph node feature vectors, using the following formula: ; in, This represents the feature weight of the graph node corresponding to the i-th superpixel. This represents the 2-norm of the feature vector of the i-th superpixel in the final image. This represents the total number of superpixels. The generated graph node feature weight matrix is ​​fed back in real time to S2's adaptive kernel regression nonlocal mean denoising algorithm, achieving adaptive optimization of the preprocessing stage.

[0052] The model training process of this algorithm is as follows: Training is performed using the publicly available vehicle scene datasets BDD100K and Cityscapes, which contain labeled semantic segmentation and object detection results. The training process employs a semi-supervised learning approach, with the pixel accuracy of semantic segmentation and the average accuracy of object detection as joint optimization objectives. The AdamW optimizer is used to iteratively update the learnable parameters. , With weighting coefficients , , The training batch size is set to 8, the learning rate is set to 0.0005, and the number of iterations is set to 300,000. After training, the optimal parameters are stored in the preset storage area of ​​the image processing domain controller for the algorithm to call in real time.

[0053] S3.6 Feature Spatiotemporal Anchoring Based on Spatiotemporal Reference. For the final graph node feature vector set, based on the spatiotemporal reference parameter package output by S1, dual spatial and temporal anchoring is completed to ensure the spatiotemporal consistency of features across the entire link. In the spatial anchoring stage, the coordinates of all superpixel centers are uniformly mapped to the vehicle body reference coordinate system to ensure that features extracted from different vehicle image acquisition nodes are in the same spatial coordinate system. In the temporal anchoring stage, all feature information is bound to the unified timestamp of the corresponding image frame. At the same time, combined with the vehicle pose data of the spatiotemporal reference unit, the inter-frame feature position offset caused by vehicle motion is compensated, and the initial predicted position for inter-frame feature matching is generated, providing a foundation for subsequent cross-frame feature verification and trajectory prediction.

[0054] S3.7 Semantic Segmentation and Object Detection Output. Based on the final graph node feature vector set, a multilayer perceptron classifier is used to complete semantic segmentation and object detection. The multilayer perceptron contains two hidden layers with dimensions of 256 and 128 respectively, and the ReLU activation function is used. The semantic segmentation stage outputs a static semantic segmentation mask, labeling the static semantic category of each superpixel, specifically including drivable areas, lane lines, curbs, traffic signs, guardrails, buildings, and trees; the object detection stage outputs a dynamic set of object detection boxes, with each detection box bound to the corresponding object category, confidence score, spatial location, and timestamp. The object categories specifically include vehicles, pedestrians, cyclists, animals, and moving obstacles.

[0055] S3.8 Feature Extraction Output Data Package Generation. This step integrates all the output results from the previous steps to generate a feature extraction output data package, which is stored in the shared memory area of ​​the image processing domain controller for real-time use in subsequent feature verification steps. The feature extraction output data package includes the final graph node feature vector set, graph node feature weight matrix, static semantic segmentation mask, dynamic object detection bounding box set, spatial anchor coordinates and temporal anchor information of all feature points, and branch feature processing weight parameters.

[0056] S4. Cross-node and cross-frame feature cross-verification and spurious feature removal This step, based on the feature extraction output data packet generated by S3, combines the preprocessing output data packet from S2 and the spatiotemporal reference parameter packet from S1 to complete the dual mutual verification of static and dynamic features, eliminating false features and invalid detection results. At the same time, it receives the false feature statistics and verification rule update parameters fed back by S8, optimizes the threshold and strictness of the verification rules, provides highly reliable feature data for subsequent semantic completion and trajectory prediction stages, and feeds back false feature-related information to subsequent stages to achieve closed-loop optimization of preprocessing and model.

[0057] S4.1 Predefined and Adaptive Adjustment of Feature Verification Rules. Based on the feature type, two types of verification rules are predefined: static feature verification rules and dynamic feature verification rules. Simultaneously, based on the regional imaging quality evaluation parameters output by S2, the stringency of the verification rules is adaptively adjusted; the worse the imaging quality, the stricter the judgment threshold of the verification rules, avoiding false feature misses caused by low-quality images. Based on the false feature statistics fed back by S8, the threshold parameters of the verification rules are updated to continuously reduce the false feature miss rate and false detection rate.

[0058] S4.2 Cross-node and cross-frame dual mutual verification of static features. For static feature points in the static semantic segmentation mask, dual verification in both spatial and temporal dimensions is performed to preemptively eliminate static pseudo-features caused by reflections, water stains, shadows, and noise, avoiding semantic map construction errors and drivable area judgment deviations caused by pseudo-features. In the cross-node spatial mutual verification stage, for static feature points within overlapping fields of view, based on the real-time extrinsic matrix set of S1, cosine similarity calculation is performed on static feature points at the same spatial location extracted from different vehicle image acquisition nodes. If the cosine similarity of feature points at the same spatial location across different nodes is not less than 0.75 and the semantic category is consistent, it is determined to be a valid static feature and retained; if the similarity is less than 0.75 or the semantic category is inconsistent, it is determined to be a feature to be verified and enters the cross-frame verification stage; for static feature points in non-overlapping fields of view, they directly enter the cross-frame verification stage. In the cross-frame temporal mutual verification step, for the static feature points of the feature points to be verified and the non-overlapping field of view, based on the spatiotemporal reference parameter package of S1 and the vehicle pose data, static feature points at the same spatial location in the preceding 5 consecutive frames of images are obtained to construct feature trajectories. If the semantic categories of the feature trajectories are continuous and consistent, and the inter-frame fluctuation of feature similarity does not exceed 10%, they are determined to be valid static features and retained; otherwise, they are determined to be static pseudo features and removed, and the region is marked as an invalid region in the static semantic segmentation mask.

[0059] S4.3 Cross-node-motion constraint dual mutual verification of dynamic features. For dynamic target detection bounding boxes, dual verification of spatial consistency and motion rationality is performed to eliminate false detections and false dynamic features in advance, avoiding trajectory prediction errors and false risk warnings caused by false detections. In the cross-node spatial mutual verification stage, for dynamic target detection bounding boxes in overlapping fields of view, based on the vehicle body reference coordinate system, the target bounding boxes detected by different vehicle image acquisition nodes are converted into three-dimensional spatial bounding boxes in the same coordinate system. If the intersection-union ratio (IU) of two three-dimensional spatial bounding boxes is not less than 0.6 and the target categories are consistent, they are determined to be valid dynamic targets. The features of the two detection boxes are fused to update the spatial position and motion parameters of the target. If the IU is less than 0.6 or the categories are inconsistent, it is determined to be a target to be verified and enters the motion constraint verification stage. For dynamic target detection bounding boxes in non-overlapping fields of view, they directly enter the motion constraint verification stage. In the motion constraint consistency verification step, for the dynamic target detection boxes of the target to be verified and the non-overlapping field of view, based on the detection results of the same target in the previous three consecutive frames of images, Kalman filtering is used to construct the target's motion trajectory and predict the target's spatial position and motion velocity in the current frame. If the lateral deviation between the detected target position and the predicted position in the current frame does not exceed 0.3 meters and the longitudinal deviation does not exceed 0.5 meters, and the motion velocity conforms to the kinematic constraints of dynamic targets in the vehicle scene, it is determined to be a valid dynamic target and retained; otherwise, it is determined to be a dynamic pseudo-feature and discarded, and the detection box is marked as invalid detection.

[0060] Examples of kinematic constraints for dynamic targets in vehicle-mounted scenarios are as follows: pedestrian speed not exceeding 15 km / h, vehicle speed not exceeding 120 km / h, cyclist speed not exceeding 40 km / h, and animal speed not exceeding 60 km / h. These can be adjusted according to the actual application scenario.

[0061] S4.4 Pseudo-feature statistics and labeling. For all removed static and dynamic pseudo-features, the occurrence area, imaging conditions, and feature type of the pseudo-features are statistically analyzed to generate pseudo-feature statistics. At the same time, in the preprocessed image data, the coordinates of the regions corresponding to the pseudo-features are labeled to generate pseudo-feature region masks, providing a basis for feedback optimization in subsequent steps.

[0062] S4.5 Feature Verification Output Data Packet Generation. This step integrates all the output results from the previous steps to generate a feature verification output data packet, which is stored in the shared memory area of ​​the image processing domain controller for real-time use in subsequent semantic completion and trajectory prediction steps. The feature verification output data packet includes a valid static feature set and a verified static semantic segmentation mask, a valid dynamic target set and a verified dynamic target detection box set, a valid detail feature map, pseudo-feature statistics, and a pseudo-feature region mask.

[0063] S5. Full-scene semantic completion and dynamic target trajectory prediction based on variational autoencoder trajectory prediction model This step, based on the effective static feature set generated in S4, the verified static semantic segmentation mask, and the effective dynamic target set, combines the spatiotemporal reference parameter package output by S1, the preprocessed image data output by S2, and the graph node feature vector set output by S3 to complete static semantic completion, dynamic multi-target tracking, and multi-temporal trajectory prediction for the entire scene. Simultaneously, it receives dynamic adaptation strategy parameters from S6 to optimize the computational granularity of trajectory prediction. The generated trajectory prediction confidence matrix is ​​fed back to the graph neural network dynamic feature interaction network in S3, achieving bidirectional collaborative optimization of feature extraction and trajectory prediction, providing core trajectory and semantic basis for subsequent scene risk identification. The variational autoencoder trajectory prediction model used in this step differs from conventional deterministic trajectory prediction algorithms. It can learn the potential multimodal distribution of the trajectory and deeply integrates static semantic environment and perceptual feature information. Its prediction accuracy and scene adaptability are significantly better than conventional algorithms, making it the core support for this solution to achieve pre-risk prediction.

[0064] S5.1 Spatiotemporal Continuous Completion of Static Semantics. For invalid regions and semantically discontinuous regions at the vehicle's field of view edges in the validated static semantic segmentation mask, two levels of spatiotemporal continuous completion are performed to resolve semantic loss issues caused by occlusion, blind spots, and pseudo-feature removal, generating a complete and continuous static semantic map. The first level is spatial completion based on the historical semantic map. Based on the effective static feature set of previous consecutive frames, a local static semantic map is constructed with the vehicle's body reference coordinate system as the reference. The local semantic map covers a 50m × 50m area around the vehicle, containing the spatial distribution information of all static semantic categories. The static semantic segmentation mask of the current frame is spatially registered with the local static semantic map. For invalid regions and semantically discontinuous regions of the current frame, the semantic information in the local static semantic map is used to complete the spatial dimension completion, generating a preliminarily completed static semantic segmentation mask. The second stage is detail completion based on semantic neighborhood constraints. For the edge regions in the static semantic segmentation mask after preliminary completion, a conditional random field model is used to combine the semantic categories of neighboring pixels with effective detail feature maps to optimize the continuity of semantic edges, eliminate semantic jaggedness and misalignment caused by completion, and generate the final full-frame static semantic segmentation result.

[0065] S5.2 Multi-target tracking and trajectory smoothing for dynamic targets. For the validated set of valid dynamic targets, a multi-target tracking algorithm based on joint probability data association, combined with Kalman filtering, is employed. A unique target ID is assigned to each valid dynamic target, and the historical motion trajectory of each target is constructed. Simultaneously, based on the target's spatial position in the vehicle's reference coordinate system, the position points in the trajectory are smoothed to eliminate trajectory jitter caused by detection noise. The real-time motion parameters of each dynamic target are updated, including 3D spatial position, speed, acceleration, and direction of motion, providing continuous and reliable historical trajectory data for subsequent trajectory prediction.

[0066] S5.3 Multi-temporal trajectory prediction based on variational autoencoder trajectory prediction model. This step uses a variational autoencoder trajectory prediction model to perform multi-temporal future trajectory prediction for each tracked dynamic target, and simultaneously generates a trajectory prediction confidence matrix, which is fed back to the graph neural network dynamic feature interaction network in S3.

[0067] The core computational process of the algorithm is as follows: The variational autoencoder trajectory prediction model consists of three parts: an encoder, a decoder, and a latent spatial distribution. First, the input vector of the encoder is constructed. The input vector integrates the target's historical motion information, semantic information of the surrounding environment, and perceptual feature information. The calculation formula is as follows: ; in, Let represent the encoder input vector for the k-th dynamic target, with dimensions set to 512. This represents the historical motion trajectory vector of the k-th dynamic target, with a dimension of 128, and is composed of the target's spatial position, velocity, and acceleration from the preceding 10 consecutive frames. The static semantic environment vector representing the k-th dynamic target is set to 128 dimensions and is composed of the feature vectors of the static semantic segmentation mask within a 30m × 30m range around the target. This represents the graph node feature vector corresponding to the k-th dynamic target, with a dimension of 256, output by the S3 graph neural network dynamic feature interaction network.

[0068] The encoder maps the input vector to the latent space and outputs the mean vector and variance vector of the latent variables, calculated as follows: ; ; in, Let represent the latent variable mean vector corresponding to the k-th dynamic objective, with a dimension of 64. This represents the learnable weight matrix for the mean mapping, with dimensions set to 64×512. Represents the learnable bias vector for the mean mapping, with a dimension of 64; Let represent the variance vector of the latent variable corresponding to the k-th dynamic objective, with a dimension of 64. This represents the learnable weight matrix for the variance mapping, with dimensions set to 64×512. This represents the learnable bias vector for the variance mapping, with a dimension of 64. This represents an exponential function used to ensure that the variance output is non-negative.

[0069] Latent variables are sampled from the latent spatial distribution. To address the gradient interruption problem during backpropagation caused by random sampling, a reparameterization technique is employed, and the calculation formula is as follows: ; in, This represents the latent variables obtained from sampling, with a dimension set to 64. This represents a random variable sampled from a standard normal distribution, with a dimension of 64. The standard normal distribution has a mean of 0 and a variance of 1.

[0070] The decoder maps the sampled latent variables to the trajectory space and outputs predicted trajectory points for future multiple time series. This example demonstrates the prediction output for the next four time series, and the calculation formula is as follows: ; in, This represents the predicted trajectory point vector of the k-th dynamic target at the τ-th time series, with a dimension of 6, containing the predicted spatial location, velocity, and direction of motion; τ represents the time series index, corresponding one-to-one with the four predicted time series: τ=1 represents the trajectory point within the next 0.5 seconds, outputting a high-confidence prediction result; τ=2 represents the trajectory point within the next 1 second, outputting a medium-to-high-confidence prediction result; τ=3 represents the trajectory point within the next 2 seconds, outputting a medium-confidence prediction result; τ=4 represents the trajectory point within the next 3 seconds, outputting a low-confidence prediction result; Let represent the learnable weight matrix for the τ-th time series, with dimensions set to 6×64. Let represent the learnable bias vector for the τth time series, with a dimension of 6.

[0071] Simultaneously, a trajectory prediction confidence matrix is ​​generated. The confidence level is calculated by combining the variance of the latent variables with the time-series weights, as shown in the following formula: ; in, This represents the prediction confidence level of the k-th dynamic target at the τ-th time series, with a value ranging from 0 to 1. This represents the adjustment factor, set to 0.1. The 2-norm represents the variance vector of the k-th dynamic objective latent variable. This represents the preset confidence weight for the τ-th time series, which corresponds one-to-one with the time series index. The value is 1.0 when τ=1, 0.8 when τ=2, 0.5 when τ=3, and 0.3 when τ=4.

[0072] The confidence score of the superpixel-level trajectory prediction corresponding to the k-th dynamic target is calculated by taking the maximum value of all time-series confidence scores, using the following formula: ; The feature weights of graph nodes corresponding to all dynamic targets are extended to all superpixels to generate a complete trajectory prediction confidence matrix. If a superpixel corresponds to multiple dynamic targets, the highest confidence score is taken; if a superpixel does not correspond to any dynamic target, the confidence score is set to 0.1. The generated trajectory prediction confidence matrix is ​​fed back to the S3 graph neural network dynamic feature interaction network in real time to dynamically adjust the graph edge weights, achieving bidirectional collaborative optimization of trajectory prediction and feature extraction.

[0073] The model training process of this algorithm is as follows: Training is performed using the publicly available vehicle scene datasets NGSIM and Argoverse, which contain labeled historical trajectories and future true trajectories of dynamic targets. The training process employs unsupervised learning, using the reconstruction loss of the variational autoencoder and the KL divergence loss as joint optimization objectives. The reconstruction loss uses the mean square error between the predicted and true trajectories, while the KL divergence loss constrains the latent spatial distribution to approximate a standard normal distribution. The learnable parameters are iteratively updated using the Adam optimizer. , , , , , The training batch size is set to 32, the learning rate is set to 0.001, and the number of iterations is set to 400,000. After training, the optimal parameters are stored in the preset storage area of ​​the image processing domain controller for the algorithm to call in real time.

[0074] In the algorithm application process, the pre-trained optimal parameters are first loaded from the preset storage area. Then, for each tracked dynamic target, the historical motion trajectory vector, static semantic environment vector, and graph node feature vector are extracted and concatenated to generate the encoder input vector. Next, the encoder maps the vector to the latent space and outputs the mean vector and variance vector. The reparameterization technique is used to sample the latent variables. Then, the decoder outputs the predicted trajectory points for the next four time series and generates the trajectory prediction confidence matrix. Finally, the predicted trajectory is constrained to a drivable area by combining the final full-frame static semantic segmentation result. If the predicted trajectory exceeds the drivable area, the confidence of the corresponding trajectory point is reduced proportionally to ensure that the prediction result conforms to the constraints of the actual scene.

[0075] S5.4 Semantic and Trajectory Output Data Package Generation. This step integrates all the output results from the previous steps to generate a semantic and trajectory output data package, which is stored in the shared memory area of ​​the image processing domain controller for real-time access in subsequent scene adaptation and risk identification stages. The semantic and trajectory output data package includes the final full-frame static semantic segmentation result, a local static semantic map, a dynamic target tracking set with unique IDs, the historical motion trajectory of each dynamic target, the predicted trajectory and corresponding confidence scores for future multi-time series, and the trajectory prediction confidence matrix.

[0076] S6. Dynamic lightweight adaptation and adaptive scheduling of computing power for the perception algorithm model This step, based on the semantic and trajectory output data packets generated by S5, combined with the feature verification output data packets output by S4, the regional imaging quality evaluation parameters output by S2, and the computing power usage data obtained in real time from the image processing domain controller system kernel and the real-time vehicle speed obtained from the vehicle CAN bus, completes the real-time assessment of scene complexity and computing power status, generates a dynamic adaptation strategy, adjusts the structure and operating parameters of the perception algorithm model, achieves a dynamic balance between model accuracy and computing power consumption, and adapts to the changing environment and computing power conditions in the vehicle scenario.

[0077] S6.1 Real-time assessment of scene complexity and computing power status. This step simultaneously performs a dual assessment of scene complexity and computing power load status, providing a basis for generating dynamic adaptation strategies. In the scene complexity assessment phase, based on the final full-frame static semantic segmentation results, the quantity and type of dynamic target tracking sets, the real-time vehicle speed, and regional imaging quality evaluation parameters, the complexity score of the current scene is calculated using the following formula: ; in, This represents the complexity score of the current scenario, ranging from 0 to 100. This represents the weight of the number of dynamic targets; the more dynamic targets there are, the higher the weight value. This represents the weight of the semantic category richness of the scene. The more complex the semantic categories and the more diverse the road environment, the higher the weight value. This indicates the weight of vehicle speed; the faster the vehicle travels, the higher the weight value. This represents the image quality weight; the worse the image quality, the higher the weight value. A higher final scene complexity score indicates a higher requirement for perception accuracy in the current scene, requiring more computing resources.

[0078] In the computing power status assessment phase, real-time computing power utilization, memory utilization, and single-frame inference latency data are obtained from the image processing domain controller to assess the current remaining computing power space, which is divided into three states: high computing power load, medium computing power load, and low computing power load. The high computing power load state refers to a state where the real-time computing power utilization is not less than 85%, the medium computing power load state refers to a state where the real-time computing power utilization is not less than 60% and less than 85%, and the low computing power load state refers to a state where the real-time computing power utilization is less than 60%.

[0079] S6.2 Dynamic adaptation strategy generation based on scenario and computing power. Based on scenario complexity scores and computing power load status, the example generates four levels of adaptation strategies, each with clearly defined model adjustment rules.

[0080] The first level is the high-precision mode. The trigger condition is that the scene complexity score is not less than 70 points and the computing power load is low or medium. The adjustment rules are to maintain the complete structure of the graph neural network dynamic feature interaction network, set the number of superpixels to 500, set the number of iterations to 3, keep all feature maps at their original resolution, set the inference accuracy to FP32, enable feature extraction and verification of the whole region, do not perform any structure or channel pruning, and prioritize the full accuracy of perception and prediction.

[0081] The second level is the balanced mode, which is triggered when the scene complexity score is not lower than 40 points and not lower than 70 points, and the computing power load is medium. The adjustment rules are as follows: the number of superpixels of the graph neural network dynamic feature interaction network is set to 400, the number of iterations is set to 2, the static environment semantic branch and the dynamic target feature branch maintain the complete structure, the number of convolution channels of the detail texture feature branch is pruned by 30%, the feature extraction resolution of non-critical regions of non-overlapping fields of view is reduced to 1 / 2 of the original resolution, and the inference accuracy is FP16. This reduces computing power consumption while ensuring the core perception accuracy.

[0082] The third level is the lightweight mode, which is triggered when the scene complexity score is below 40 and the computing power load is medium or high. The adjustment rules are as follows: the number of superpixels of the graph neural network dynamic feature interaction network is set to 300, the number of iterations is set to 1, the static environment semantic branch only retains the semantic category output related to the drivable area and lane lines, the dynamic target feature branch only retains the detection of three core targets: vehicles, pedestrians and cyclists, the number of convolution channels of the detail texture feature branch is pruned by 50%, and feature extraction is only performed on the overlapping field of view and the area where the dynamic target is located. Feature extraction of non-critical areas is turned off, the inference accuracy is INT8, and the perception efficiency of the core scene is prioritized to reduce computing power consumption.

[0083] Level 4 is the emergency mode, triggered when the computing load is high and the scene complexity score is not lower than 60. The adjustment rules are as follows: the number of superpixels of the graph neural network dynamic feature interaction network is set to 250, the number of iterations is set to 1, unnecessary calculations of the detail texture feature branch are turned off, the static environment semantic branch only outputs the mask of the drivable area, the dynamic target feature branch only performs full-precision detection on targets within a 120-degree field of view in front of the vehicle, the detection resolution of the other field of view areas is reduced to 1 / 4 of the original resolution, the inference accuracy is INT8, and priority is given to ensuring the perception accuracy and inference speed of the core area of ​​vehicle driving to cope with high-load and high-risk scenarios.

[0084] S6.3 Adaptation Strategy Execution and Parameter Update. The generated dynamic adaptation strategy parameters, including the number of superpixels, iteration count, branch pruning rules, resolution adjustment parameters, inference accuracy settings, and feature extraction region range, are updated to the control module of the perceptual algorithm model and fed back to the feature extraction stage in S3. During feature extraction of the next frame, the corresponding adaptation strategy is executed. Simultaneously, based on the adaptation strategy, the semantic and trajectory information of non-critical regions is simplified based on the processing results of the current frame, retaining only the core effective information to reduce the computational burden on subsequent stages.

[0085] S6.4 Model Adaptation Output Data Package Generation. This step integrates the output results of the previous steps to generate a model adaptation output data package, which is stored in the shared memory area of ​​the image processing domain controller for subsequent risk identification. Simultaneously, the dynamic adaptation strategy parameters are fed back to the S3 step in real time. The model adaptation output data package includes scene complexity score, computing load status, dynamic adaptation strategy parameters, and simplified core semantics and target information.

[0086] S7. Risk Scene Classification and Identification Based on Semantics and Trajectory, and Targeted Image Enhancement Output This step, based on the semantic and trajectory output data packets generated by S5, the scene complexity score and core semantic target information output by S6, combined with the spatiotemporal reference parameter packets output by S1 and the real-time vehicle body status data obtained from the vehicle CAN bus, completes the real-time identification and level determination of risk scenarios. Based on the risk level, it executes differentiated image enhancement and output strategies, and feeds back the risk area coordinates to the preprocessing stage of S2 to realize the pre-enhancement processing of risk areas, providing intuitive and high-priority image and risk information output for the driver and autonomous driving decision-making system.

[0087] S7.1 Risk Scenario Grading Rules Predefined. Based on the degree of impact on vehicle driving safety, four risk levels are exemplified, each with a specific trigger condition. Safety level 0 is triggered when there are no dynamic targets around the vehicle, or the predicted trajectory of a dynamic target does not intersect with the vehicle's own trajectory, the drivable area is continuous and unobstructed, and there is no risk of traffic sign violations. Warning level 1 is triggered when there are dynamic targets around the vehicle, but the longitudinal distance between the target and the vehicle is not less than 30 meters, the lateral distance is not less than 5 meters, the predicted trajectories do not intersect, or the vehicle shows a tendency to deviate from the lane line but does not exceed the lane line, or there are warning traffic signs such as speed limits or yield signs. Alert level 2 is triggered when a dynamic target and the vehicle... The longitudinal distance between the target and the vehicle is not less than 10 meters and less than 30 meters, and the lateral distance is not less than 1 meter and less than 5 meters. The predicted trajectory has a potential intersection, or the vehicle has deviated from the lane line, or there is a static obstacle close to the vehicle, or there is a risk of running a red light or speeding. The emergency level is 3. The triggering conditions are that the longitudinal distance between the dynamic target and the vehicle is less than 10 meters and the lateral distance is less than 1 meter. The predicted trajectory has a direct collision risk, or a sudden obstacle appears in front of the vehicle and the collision time is less than 2 seconds, or there are emergency scenarios such as the lane line suddenly disappearing or the drivable area being interrupted.

[0088] S7.2 Real-time Risk Scene Identification and Level Determination. Based on real-time vehicle body state data, including vehicle speed, steering wheel angle, throttle opening, brake signal, and turn signal, a vehicle dynamics model is used to predict the vehicle's trajectory within the next 3 seconds. Combining the dynamic target prediction trajectory output from S5 and the full-frame static semantic segmentation results, the collision time and minimum safe distance between the vehicle's own trajectory and the dynamic target trajectory are calculated. Simultaneously, risk factors such as static obstacles, traffic signs, and lane line status are identified. Based on predefined grading rules, the risk level of the current scene is determined, and the image region coordinates corresponding to all risk factors are located, generating a risk region coordinate set.

[0089] S7.3 Targeted Image Enhancement and Output Based on Risk Level. Differentiated image processing and output strategies are implemented for different risk levels, covering all risk levels to ensure effective transmission of risk information. For Level 0 safety, globally stitched surround view and forward main view images are output, maintaining the standard pre-processed effect without additional enhancement, while simultaneously outputting risk-free structured semantic information. For Level 1 warning, risk areas are highlighted in the output global image, with slight contrast enhancement to increase driver attention, and warning-type semantic information is output simultaneously. For Level 2 alert, locally magnified images of the risk areas are prioritized for output, with a magnification of 2x, and contrast enhancement and edge sharpening are applied to the magnified areas. The risk areas are also marked with red borders in the global image, and warning information and risk factor details are output simultaneously. For Level 3 emergency, magnified images of the risk areas are output in full screen, with a magnification of 3x, and extreme contrast enhancement, deblurring, and edge strengthening are applied. Collision warning signs are overlaid, and emergency braking and avoidance prompts are output simultaneously, all completed with the highest priority.

[0090] S7.4 Risk and Image Output Data Packet Generation. This step integrates the output results of the previous steps to generate risk output and image output data packets. These packets are transmitted to the vehicle's human-machine interface system and autonomous driving decision-making system for execution, and stored in the shared memory area of ​​the image processing domain controller for subsequent closed-loop verification. Simultaneously, the risk area coordinate set is fed back to the preprocessing stage in S2 for targeted region enhancement in the next frame of imagery, achieving pre-optimization of risk areas. The risk output and image output data packets include the scene risk level, risk factor details, risk area coordinate set, graded enhanced output image data, and structured output information on semantics and targets.

[0091] S8. Closed-loop verification of end-to-end processing results and incremental self-calibration of the sensing algorithm model. This step, based on the spatiotemporal reference parameter package of S1, the pseudo-feature statistical information of S4, the semantic and trajectory results of S5, and the risk output data of S7, combined with the true reference data obtained from the vehicle-mounted LiDAR perception system and the actual vehicle driving data obtained from the spatiotemporal reference unit, completes the closed-loop verification of the end-to-end processing results, generates corresponding calibration signals and update parameters, and feeds them back to the previous steps, realizing the self-optimization and long-term stability improvement of the entire method, forming a complete closed-loop architecture of "processing-verification-feedback-optimization".

[0092] S8.1 Closed-loop verification and calibration triggering of spatiotemporal reference parameters. Based on the ground truth reference data of LiDAR, the reprojection error is calculated for the spatial coordinates of static feature points in the static semantic segmentation results. If the average reprojection error is not less than 0.5 pixels, it is determined that there is a drift in the spatiotemporal reference parameters. An extrinsic or intrinsic calibration trigger signal is generated and fed back to the parameter update callback mechanism of S1, triggering S1 to re-execute the dynamic calibration of intrinsic and extrinsic parameters, update the spatiotemporal reference parameter package, and solve the spatial positioning error problem of the entire link from the source.

[0093] S8.2 Error Feedback and Parameter Update in the Preprocessing Stage. Based on the pseudo-feature statistics output from S4, the regions and imaging conditions where pseudo-features are concentrated are statistically analyzed. If the frequency of pseudo-features in a certain type of region or under a certain imaging condition is not less than 30%, a preprocessing parameter update signal is generated and fed back to the normalization processing stage in S2. The normalization and enhancement processing parameters for the corresponding region and imaging condition are adjusted to improve the image preprocessing quality and reduce the generation of pseudo-features. At the same time, the pseudo-feature region mask is fed back to S2 for targeted enhancement processing in the next frame, reducing the probability of pseudo-features from the preprocessing stage.

[0094] S8.3 Feature Error Analysis and Incremental Self-calibration of the Perception Algorithm Model. First, feature error analysis is performed by comparing the validated effective static feature set, full-frame static semantic segmentation results, and effective dynamic target set with the ground truth reference data from the LiDAR. This involves statistically analyzing misclassified pixels in semantic segmentation and false positives or false negatives in target detection, identifying the corresponding feature types, feature regions, and imaging scenes, and generating model error statistics. Next, incremental self-calibration is performed. Based on the model error statistics, image segments and feature data corresponding to the errors are extracted to construct an incremental training sample set. A low-rank adaptive incremental learning method is used to fine-tune the head network of the graph neural network's dynamic feature interaction network, updating only the weight parameters of the head network without modifying the weights of the backbone network to avoid catastrophic forgetting. After fine-tuning, the updated weight parameters of the perception algorithm model are generated and fed back to the model initialization stage in S3. In the processing of the next frame, the updated model weights are used to continuously improve the accuracy and scene adaptability of the perception model.

[0095] Quantitative evaluation of the S8.4 end-to-end processing performance. For each frame's processing result, core performance metrics of the entire process are calculated, including pixel accuracy and class intersection-over-union ratio for static semantic segmentation, precision, recall, and mean precision for dynamic object detection, ID switching rate and trajectory accuracy for object tracking, and end-to-end latency for image processing. All performance metrics are recorded in the log storage area of ​​the image processing domain controller for long-term performance monitoring and optimization analysis.

[0096] S8.5 Closed-loop calibration signal output. This step integrates all calibration and update information from the preceding steps, generates a closed-loop calibration output data package, and feeds back the corresponding calibration trigger signal, parameter update signal, and model weight update parameters to the corresponding steps S1, S2, and S3, respectively, to complete the closed-loop update and self-optimization of the entire chain, ensuring the long-term stability, accuracy, and robustness of this method in complex automotive scenarios.

[0097] The automotive intelligent image processing system based on the perception algorithm model, applying the automotive intelligent image processing method based on the perception algorithm model described above, comprises: a spatiotemporal reference calibration module for performing operation S1 and generating a spatiotemporal reference parameter package; an adaptive preprocessing module for performing operation S2 and generating a preprocessing output data package; a graph neural network feature extraction module for performing operation S3 and generating a feature extraction output data package; a pseudo-feature verification and removal module for performing operation S4 and generating a verified valid feature data package; a semantic completion and trajectory prediction module for performing operation S5 and generating semantic and trajectory output data packages; a computing power adaptive scheduling module for performing operation S6 and generating dynamic adaptation strategy parameters; a risk classification and image output module for performing operation S7 and completing risk identification and image enhancement output; and a full-link closed-loop calibration module for performing operation S8 and completing full-link closed-loop verification and self-optimization.

Claims

1. A vehicle intelligent image processing method based on a perception algorithm model, characterized in that, include: S1. Establish a globally unified time reference and vehicle body reference coordinate system, complete the dynamic calibration of intrinsic and extrinsic parameters corresponding to image acquisition, and generate a unified spatiotemporal reference parameter package for the entire link; S2. Based on the spatiotemporal reference parameter package, the vehicle-mounted raw image data is sequentially processed for distortion correction, cross-node field of view division, heterogeneous domain normalization and adaptive denoising, generating a preprocessed output data package; S3. Based on the spatiotemporal reference parameter package and the preprocessed output data package, image hierarchical feature extraction is completed through the graph neural network dynamic feature interaction network. Static semantic segmentation and dynamic target detection are completed based on the extracted features, and feature extraction output data package is generated. S4. Based on the spatiotemporal reference parameter package, preprocessed output data package and feature extraction output data package, perform double cross-verification on static features and dynamic features respectively, eliminate false features, and generate valid feature data packages after verification. S5. Based on the spatiotemporal reference parameter package, feature extraction output data package and effective feature data package, complete the spatiotemporal continuous completion of static semantics and multi-target tracking of dynamic targets. Through the variational autoencoder trajectory prediction model, complete the multi-temporal trajectory prediction of dynamic targets and generate semantic and trajectory output data packages. S6. Based on semantic and trajectory output data packets, combined with real-time computing power usage data and vehicle real-time driving speed, complete the scene complexity and computing power load status assessment, and generate dynamic adaptation strategy parameters; S7. Based on semantic and trajectory output data packets and dynamic adaptation strategy parameters, combined with real-time vehicle body status data, complete the risk scene classification identification, and perform differentiated image enhancement and output based on the risk level.

2. The intelligent image processing method for automobiles based on a perception algorithm model according to claim 1, characterized in that, Also includes: S8. Based on external truth reference data and combined with the output results of each step in the entire link, complete the closed-loop verification of the entire link processing results, generate corresponding calibration and update parameters, and complete the self-optimization of the entire link; the closed-loop verification includes three categories: closed-loop verification of spatiotemporal reference parameters, closed-loop optimization of preprocessing parameters, and incremental self-calibration of the perception algorithm model; The spatiotemporal reference parameter closed-loop verification calculates the reprojection error of static feature points based on true reference data. When the error exceeds the preset threshold, a calibration trigger signal is generated and fed back to S1. The preprocessing parameter closed-loop optimization is based on pseudo-feature statistical information, and the generated parameter update signal is fed back to S2; the perception algorithm model incremental self-calibration is based on true reference data to complete feature error analysis, construct an incremental training sample set to complete model fine-tuning, and the generated model weight update parameters are fed back to S3.

3. The intelligent image processing method for automobiles based on a perception algorithm model according to claim 1, characterized in that, In S1, the vehicle body reference coordinate system is a three-dimensional rectangular coordinate system established with the rear axle center of the vehicle as the origin, the vehicle's driving direction as the positive X-axis, the vehicle's vertical upward direction as the positive Z-axis, and the vehicle's left lateral direction as the positive Y-axis. The global unified time reference is established by accurately marking the global unified timestamp onto the frame header of each frame of the original vehicle image data to ensure that the time synchronization error of all image data does not exceed 1 millisecond. The dynamic calibration of intrinsic parameters is completed based on the static feature constraints of the natural scene and the bundle adjustment method, while the dynamic calibration of extrinsic parameters is completed based on the vehicle pose recursion and cross-node feature matching.

4. The intelligent image processing method for automobiles based on a perception algorithm model according to claim 1, characterized in that, In S2, the adaptive denoising is performed using an adaptive kernel regression nonlocal mean denoising algorithm. This algorithm simultaneously integrates the spatial distance constraints of local pixels and the gray-level similarity constraints of nonlocal pixels to calculate the denoising weights for each pixel, thus completing image denoising. The denoising weights are adaptively adjusted based on the graph node feature weight matrix generated in S3. The denoising intensity is reduced in areas with high feature weights to preserve detailed features, while the denoising intensity is increased in areas with low feature weights to suppress noise interference.

5. The intelligent image processing method for automobiles based on a perception algorithm model according to claim 1, characterized in that, In S3, the execution process of the graph neural network dynamic feature interaction network is as follows: superpixels are generated from the preprocessed image data, an initial dynamic graph structure is constructed with each superpixel as a graph node, the initial feature vector of each graph node is extracted, the initial graph edge weights are calculated, and multiple rounds of graph node feature iteration updates are completed through the graph attention mechanism to generate the final graph node feature vector set and graph node feature weight matrix; the initial graph edge weights are dynamically adjusted based on the trajectory prediction confidence matrix generated in S5 to improve the feature interaction intensity of graph nodes corresponding to high-confidence dynamic targets.

6. The intelligent image processing method for automobiles based on a perception algorithm model according to claim 1, characterized in that, In S4, the dual mutual verification performed on static features is cross-node spatial mutual verification and cross-frame temporal mutual verification, which determines the validity of static features based on feature similarity, semantic category consistency and inter-frame feature trajectory stability; the dual mutual verification performed on dynamic features is cross-node spatial mutual verification and motion constraint consistency verification, which determines the validity of dynamic features based on spatial position consistency, target category consistency and vehicle scene kinematic constraints. The pseudo-feature region mask generated during the verification process is fed back to S2 for targeted region enhancement processing in subsequent image frames.

7. The intelligent image processing method for automobiles based on a perception algorithm model according to claim 1, characterized in that, In S5, the execution process of the variational autoencoder trajectory prediction model is as follows: the historical motion trajectory information of the dynamic target, the surrounding static semantic environment information, and the corresponding graph node feature information are fused to construct the encoder input vector. The encoder maps the input vector to the latent space, and the latent variables are sampled using the reparameterization technique. Then, the decoder outputs the predicted trajectory points of the dynamic target in the future multi-time series, and generates the trajectory prediction confidence matrix. The trajectory prediction confidence matrix is ​​fed back to S3 to dynamically adjust the graph edge weights of the dynamic feature interaction network of the graph neural network.

8. The intelligent image processing method for automobiles based on a perception algorithm model according to claim 1, characterized in that, In S6, the scene complexity score is calculated by weighting the dynamic target quantity weight, scene semantic category richness weight, vehicle speed weight, and image imaging quality weight. The computing power load state is divided into three categories: high computing power load state, medium computing power load state, and low computing power load state. Based on the scene complexity score and computing power load state, corresponding dynamic adaptation strategy parameters are generated. The dynamic adaptation strategy parameters are fed back to S3 to adjust the graph neural network structure, feature extraction rules, and inference accuracy of subsequent image frames.

9. The intelligent image processing method for automobiles based on a perception algorithm model according to claim 1, characterized in that, In S7, four risk levels are predefined: security level, alert level, warning level, and emergency level. Each risk level corresponds to a preset trigger condition. The risk level determination process is as follows: predict the vehicle's future driving trajectory based on real-time vehicle body status data, calculate the collision time and minimum safe distance by combining the predicted trajectory of dynamic targets, identify risk factors by combining static semantic information, and complete the risk level determination. The higher the risk level, the higher the magnification and enhancement intensity of the corresponding risk area, and the higher the output priority. The identified risk area coordinate set is fed back to S2 for targeted area enhancement processing in subsequent image frames.

10. An intelligent image processing system for automobiles based on a perception algorithm model, characterized in that, The automotive intelligent image processing method based on a perception algorithm model, as described in any one of claims 2 to 9, wherein the automotive intelligent image processing system based on the perception algorithm model comprises: The spatiotemporal reference calibration module is used to perform the operation of S1 and generate a spatiotemporal reference parameter package; The adaptive preprocessing module is used to perform the S2 operation and generate preprocessed output data packets; The graph neural network feature extraction module is used to perform the operation of S3 and generate feature extraction output data packets; The pseudo-feature verification and elimination module is used to perform the S4 operation and generate a valid feature data packet after verification. The semantic completion and trajectory prediction module is used to perform the operations of S5 and generate semantic and trajectory output data packets; The computing power adaptive scheduling module is used to execute S6 operations and generate dynamically adaptable strategy parameters; The risk classification and image output module is used to perform the operations of S7, completing risk identification and image enhancement output; The end-to-end closed-loop calibration module is used to perform S8 operations and complete end-to-end closed-loop verification and self-optimization.

Images