Snowfield trajectory and extreme environment intelligent identification method based on AI

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to an AI-based snowfield trajectory and extreme environment intelligent identification method, which comprises the following steps: a multi-modal data acquisition module; a fusion type deep learning model; a confrontation generation data enhancement unit; an adaptive control subsystem; and an edge calculation node deploys a lightweight model inference engine to realize low-delay trajectory feature extraction and classification decision. According to the method, a multi-scale feature pyramid structure and a cross-modal attention mechanism are adopted, the convolutional neural network and the recurrent neural network are combined, and spatial features and modeling time sequence dynamics can be extracted at the same time. The synthetic data of the extreme snowfield scene is generated by using the conditional generative adversarial network, and the physical constraint generator and the multi-scale discriminator are combined, so that the real snow particle motion trail can be simulated, and the global consistency and the local texture authenticity of the generated data are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to an intelligent recognition method for snow tracks and extreme environments based on AI. Background Art

[0002] Traditional snow track recognition methods mainly rely on single visible light cameras or artificial feature extraction technologies, and there are significant technical limitations in complex and changeable snow environments, which are specifically manifested as the following three defects:

[0003] First of all, under extreme environmental conditions, the recognition performance of traditional methods drops sharply. When encountering blizzard weather, snowflakes will cause serious scattering and occlusion effects on visible light, resulting in a significant reduction in image contrast; in low light conditions (such as at night or in polar night environments), visible light imaging systems are difficult to obtain sufficient scene information, and the image signal-to-noise ratio drops significantly. Research shows that under blizzard conditions, the recognition accuracy of traditional visible light cameras may drop by more than 60%.

[0004] Secondly, the low texture characteristics of the snow surface pose a huge challenge to feature extraction. The snow-covered ground often presents large areas of uniform white regions, lacking obvious texture features and edge information. Artificial feature extraction methods (such as SIFT, SURF, etc.) relied on by traditional algorithms are difficult to find enough feature points for matching in such an environment, resulting in a significant increase in the recognition failure rate. Experimental data shows that in flat snow areas, the number of effective feature points of traditional feature extraction algorithms may be reduced by more than 80%.

[0005] Finally, the motion characteristics of the mobile platform seriously affect the quality of data collection. When the device carried is driving on the snow, platform vibrations will cause image blurring and jitter, especially under conditions of high-speed movement or rough terrain, this phenomenon is more obvious. At the same time, rapid movement may lead to insufficient image acquisition frame rate, resulting in the loss of key feature information. Field tests show that when the moving speed exceeds 15 km / h, the track recognition accuracy of traditional methods will drop by more than 40%.

[0006] These technical defects seriously restrict the actual application effects of traditional snow track recognition methods in fields such as polar scientific research, snow rescue, and winter military operations, and it is urgent to break through the existing technical bottlenecks through new technical means such as multi-modal perception and deep learning. Summary of the Invention

[0007] Aiming at the technical problems of sharp drop in recognition rate due to poor image quality in the above extreme environment, sparse features in low-texture areas of snow making traditional algorithms easily fail, and insufficient data collection stability when the mobile platform vibrates or moves too fast, the present invention provides an intelligent recognition method for snow tracks and extreme environments based on AI.

[0008] To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0009] An AI-based intelligent recognition method for snow tracks and extreme environments, comprising the following steps:

[0010] S1. A multi-modal data acquisition module, including a high-definition camera, an infrared camera, and a lidar, is used to synchronously obtain visible light images, infrared thermal images, and three-dimensional point cloud data of the snow environment;

[0011] S2. A fusion deep learning model extracts spatial features by a convolutional neural network (CNN), models temporal dynamics by a recurrent neural network (RNN), and fuses multi-source data features through a cross-modal attention mechanism;

[0012] S3. An adversarial generated data augmentation unit uses a conditional generative adversarial network (cGAN) to generate synthetic data of extreme snow scenes, including virtual track samples under blizzards, reflective interference, and low contrast conditions;

[0013] S4. An adaptive control subsystem integrates an inertial measurement unit (IMU) and a road surface adhesion coefficient estimator to dynamically adjust the driving speed, suspension height, and sensor sampling frequency of the mobile platform;

[0014] S5. An edge computing node deploys a lightweight model inference engine to achieve low-latency trajectory feature extraction and classification decision-making.

[0015] The multi-modal data acquisition module in S1 includes: a polarized light imaging unit for suppressing snow mirror reflection interference; a multi-baseline stereo vision array configured with a camera group with different field of view angles to achieve wide-range coverage; an automatic calibration device for lidar and visible light images to achieve sub-pixel level spatial alignment through a checkerboard target.

[0016] The fusion deep learning model in S2 adopts a multi-scale feature pyramid structure, including: parallel residual CNN branches for processing different resolution image inputs; a bidirectional gated recurrent unit (Bi-GRU) for capturing trajectory continuity features; a feature fusion layer for achieving cross-modal feature alignment through learnable weights, and the weight coefficients are dynamically adjusted by environmental visibility parameters.

[0017] The training method of the fusion deep learning model in S2 is as follows:

[0018] S2.1. Adopt a curriculum learning strategy to train the model in stages according to the severity of the environment;

[0019] S2.2. Introduce a contrast loss function to enhance the tightness of similar trajectory features;

[0020] S2.3. Use differentiable data augmentation technology to achieve end-to-end robustness optimization.

[0021] The adversarial generative data augmentation unit in S3 includes: a physical constraint generator that integrates a fluid dynamics model to simulate the movement trajectory of real snow grains; a discriminator network that adopts a multi-scale PatchGAN architecture to evaluate both the global consistency and local texture authenticity of the image; and a data augmentation strategy controller that automatically adjusts the difficulty level of the generated samples according to the model training error distribution.

[0022] The adaptive control subsystem in S4 includes: a vibration suppression algorithm based on road surface unevenness spectrum estimation that adjusts the stability of the pan-tilt through feedforward-feedback composite control; a dynamic exposure control module that optimizes the shutter speed and ISO parameters of the camera in real time according to the snow reflectivity; and a lidar scanning mode switching mechanism that automatically enables the high-density scanning mode when the visibility is below the threshold.

[0023] The edge computing node in S5 adopts: model distillation technology to transfer the decision-making knowledge of the teacher network to the lightweight student network; a dynamic computing offloading mechanism that distributes computing tasks among the edge and cloud according to the network bandwidth; and a data caching strategy based on spatio-temporal correlation that preferentially retains the original data of abnormal trajectory segments.

[0024] The adaptive control subsystem integrates an abnormal trajectory warning module, and its implementation method is as follows:

[0025] S4.1. Construct a trajectory dynamics feature template library, including typical abnormal patterns such as vehicle skidding and pedestrian falling.

[0026] S4.2. Design a spatio-temporal graph convolutional network (ST-GCN) to detect abnormal trajectory continuity.

[0027] S4.3. Integrate the confidence scores of multiple sensors to generate a comprehensive warning level.

[0028] A system for an AI-based intelligent recognition method of snow tracks and extreme environments, and the application method of the system on a mobile platform includes:

[0029] Step 1. Establish a mapping table of snow adhesion coefficient - safe speed, and dynamically optimize path planning through reinforcement learning.

[0030] Step 2. Deploy a tightly coupled visual-inertial SLAM system to achieve precise positioning in a GPS-free environment.

[0031] Step 3. Design a modular thermal management device to ensure the working stability of each sensor under extreme low temperatures.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] 1. By integrating multi-modal data acquisition modules such as high-definition cameras, infrared cameras, lidar, and polarization imaging units, the present invention can synchronously obtain visible light images, infrared thermal images, and three-dimensional point cloud data of the snow environment, and effectively suppress the interference of snow mirror reflection. Combining a multi-baseline stereo vision array and an automatic calibration device, wide-range coverage and sub-pixel level spatial alignment are achieved, significantly improving the accuracy and robustness of trajectory recognition in complex snow environments.

[0034] 2. The present invention adopts a multi-scale feature pyramid structure and a cross-modal attention mechanism, combines convolutional neural network (CNN) and recurrent neural network (RNN), and can simultaneously extract spatial features and model temporal dynamics. Through the curriculum learning strategy and contrastive loss function, the model can adapt to different harsh environments in stages, enhance the tightness of similar trajectory features, and thus maintain high recognition performance under extreme conditions (such as blizzards and low-contrast environments).

[0035] 3. The present invention uses a conditional generative adversarial network (cGAN) to generate synthetic data of extreme snow scenes. Combining a physical constraint generator and a multi-scale PatchGAN discriminator, it can simulate the movement trajectories of real snow grains and ensure the global consistency and local texture authenticity of the generated data. The data augmentation strategy controller dynamically adjusts the difficulty of the generated samples according to the training error, further improving the generalization ability of the model in extreme environments.

[0036] 4. The integrated inertial measurement unit (IMU) and road surface adhesion coefficient estimator of the present invention can real-time adjust the driving speed, suspension height, and sensor parameters of the mobile platform through vibration suppression algorithms, dynamic exposure control, and lidar scan mode switching mechanisms, ensuring the stable operation of the system under complex road surfaces and harsh weather conditions.

[0037] 5. The present invention adopts model distillation technology and a dynamic computing offloading mechanism, transfers the decision-making knowledge of the teacher network to the lightweight student network, and distributes computing tasks among the edge and cloud according to the network bandwidth, significantly reducing the computing latency. The data caching strategy based on spatio-temporal correlation preferentially retains abnormal trajectory segment data, providing high-quality data support for subsequent analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, other implementation drawings can be obtained by extension based on the provided drawings without creative efforts.

[0039] The structures, proportions, sizes, etc. illustrated in this specification are only used to cooperate with the content disclosed in the specification for those familiar with this technology to understand and read, and are not used to limit the conditions for the implementation of the present invention. Therefore, they do not have substantial technical significance. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope covered by the technical content disclosed in the present invention.

[0040] Figure 1 It is a flowchart of the steps of the present invention. Specific embodiments

[0041] To make the purposes, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. These descriptions are only to further illustrate the features and advantages of the present invention, rather than a limitation on the claims of the present invention; based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present application.

[0042] The following will further describe in detail the specific embodiments of the present invention in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0043] The AI-based snow track and extreme environment intelligent recognition method and system of the present invention achieve high-precision track recognition and safety control in complex snow scenes through multi-modal data fusion, deep learning modeling, dynamic data enhancement, and adaptive control. As Figure 1 shown, the following will detail the implementation steps and technical details in conjunction with specific embodiments.

[0044] I. Multi-modal data acquisition and preprocessing

[0045] 1. Polarized light imaging unit

[0046] A split focal plane polarization camera (such as Sony IMX250MYR) is used to suppress the specular reflection of snow through Stokes vector calculation. A four-way polarization filter (0°, 45°, 90°, 135°) is set, and the polarization degree (DoP) is used to separate the diffuse reflection and specular reflection components to improve the visibility of low-contrast snow tracks.

[0047] 2. Multi-baseline stereo vision array

[0048] Deploy three groups of cameras (wide - angle 120°, medium - focal 60°, long - focal 30°) to form a stereo vision system with baseline lengths of 0.5m, 1.2m, and 2m respectively. Generate multi - resolution depth maps through a disparity fusion algorithm, covering a detection range from 5m to 50m, and adapting to the synchronous detection of short - range obstacles and long - range trajectories in the snow.

[0049] 3. LiDAR - Visible Light Automatic Calibration

[0050] Use a checkerboard target (black - and - white square size 10mm×10mm) for joint calibration. Align the LiDAR point cloud and visible - light image through the Iterative Closest Point (ICP) algorithm to achieve sub - pixel - level spatial alignment (error < 0.2 pixels). After calibration, the scan data of the LiDAR (such as Velodyne VLP - 16) and the RGB image are synchronized through time - space stamps to construct a multi - modal training dataset.

[0051] II. Construction and Training of the Fusion - type Deep - Learning Model

[0052] 1. Network Architecture Design

[0053] Multi - scale CNN branches: Adopt a parallel structure of ResNet - 50 and MobileNetV3 to process inputs with resolutions of 1080p and 480p respectively, and extract local texture and global contour features.

[0054] Temporal modeling module: A bidirectional Bi - GRU network (hidden layer with 256 dimensions) processes a sequence of 10 consecutive frames of images to capture the continuity of trajectory movement.

[0055] Cross - modal attention fusion: Design a learnable weight matrix to align CNN and RNN features, and the weight coefficients are adjusted in real - time by an environmental visibility sensor (for example, when the visibility < 50m, the RNN weight is increased by 30%).

[0056] 2. Model Training Strategy

[0057] Curriculum learning: Train in three stages - ① sunny snow scenes (visibility > 1km); ② moderate - snow scenes (visibility 200 - 500m); ③ blizzard scenes (visibility < 50m).

[0058] Contrastive loss function: Construct a triplet loss (margin = 0.5) in the feature space to enhance the feature compactness of similar trajectories (such as vehicle tire tracks), and expand the distance between dissimilar trajectories (animal footprints) by 20%.

[0059] Differentiable data augmentation: Randomly apply snow - fog simulation (random scattering coefficient β∈

[0060] [0.01, 0.1]) and motion blur (kernel size 5×5) during training to achieve end - to - end robustness optimization.

[0061] III. Adversarial Generated Data Augmentation

[0062] 1. Conditional Generative Adversarial Network (cGAN) Structure

[0063] Generator: The U-Net architecture integrates a physical constraint layer, simulates the movement trajectory of snow grains through the Navier-Stokes equation, and generates synthetic images with adjustable blizzard intensity (wind speed 5 - 15 m / s).

[0064] Discriminator: Multi-scale PatchGAN (scales 64×64, 128×128, 256×256) evaluates the authenticity of images, and the local Patch loss weight accounts for 60%.

[0065] Difficulty Adaptation: Monitor the F1-score of the model on the validation set. When the accuracy > 85%, the generator increases the specular interference (reflectivity increased to 70% - 90%) and the motion blur intensity.

[0066] 2. Synthetic Data Validation

[0067] The generated samples and real data are evaluated for similarity through the Frechet Inception Distance (FID < 25), and manual annotations of trajectory key points (such as inflection points, braking marks) are added for verification to ensure the physical rationality of the synthetic data.

[0068] IV. Adaptive Control Subsystem

[0069] 1. Road Surface Adhesion Coefficient Estimation

[0070] Based on the longitudinal acceleration from the IMU (sampling rate 200 Hz) and the wheel speed sensor data, the snow friction coefficient μ (range 0.1 - 0.3) is calculated in real time through the Extended Kalman Filter (EKF), and a μ - safe speed mapping table is constructed (e.g., when μ = 0.2, the speed limit is 30 km / h).

[0071] 2. Dynamic Parameter Adjustment

[0072] Suspension Height Control: Use a PID controller to adjust the hydraulic suspension. When the road surface unevenness > 0.1 m, raise the chassis height to 300 mm.

[0073] Sensor Optimization: Automatically reduce the camera exposure time to 1 / 1000 s according to the snow reflectivity (> 5000 lux), and switch the lidar to the high-density mode (scan lines 32 lines → 64 lines) when the visibility < 100 m.

[0074] 3. Abnormal Trajectory Warning

[0075] ST-GCN Detection: The spatio-temporal graph convolutional network extracts the motion features of trajectory nodes (sampling interval 0.5 s) and detects anomalies such as sudden acceleration changes (> 3 m / s 2 ) or heading angle deviation (> 15°).

[0076] Multi-Sensor Confidence Fusion: Set the confidence weights of lidar, camera, and IMU to 0.6, 0.3, and 0.1 respectively. When the comprehensive score is lower than 0.7, a level-3 warning is triggered.

[0077] V. Edge Computing Node Deployment

[0078] 1. Lightweight Inference Engine

[0079] Adopt model distillation technology to transfer the knowledge of the teacher network (ResNet-50) to the student network (MobileNetV3), compress the model size to 1 / 4 (from 85 MB → 21 MB), and the inference latency < 50 ms (NVIDIA Jetson Xavier platform).

[0080] 2. Dynamic Computing Offloading

[0081] Allocate tasks according to the bandwidth status: When the bandwidth > 50 Mbps, the original data is uploaded to the cloud for training; when the bandwidth < 10 Mbps, the edge node only performs key feature extraction (such as trajectory curvature and length).

[0082] 3. Data Caching Strategy

[0083] Adopt the LRU caching mechanism to preferentially retain the original data of abnormal trajectory segments (such as detecting slippage for 3 consecutive frames), and extend the caching duration to 30 minutes for post-event analysis.

[0084] VI. System Integration and Application

[0085] 1. Snowy Path Planning

[0086] Dynamically optimize the path based on the Q-learning algorithm. The reward function considers the adhesion coefficient (weight 0.6), energy consumption (weight 0.3), and trajectory smoothness (weight 0.1). After 1000 iterations, it converges to the optimal path.

[0087] 2. GPS-Free Positioning

[0088] A vision-inertial tightly coupled SLAM system that fuses IMU pre-integration and visual feature points through the ORB-SLAM3 framework, with a positioning error < 0.5 m (tested on a 1 km trajectory).

[0089] 3. Low-Temperature Thermal Management

[0090] Design a modular heating device (ceramic heating element + PID temperature control) to maintain the sensor operating in an environment of -40°C and keep the temperature of the camera lens > 0°C to avoid frosting.

[0091] Through the above implementation, the system can achieve a trajectory recognition accuracy of ≥ 92.5% and a warning response delay of < 200 ms in extreme snowy environments such as blizzards and strong reflections, significantly improving the safety and reliability of the mobile platform.

[0092] The above only elaborates on the preferred embodiments of the present invention in detail. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention, and all such changes should be included within the protection scope of the present invention.

Claims

1. An AI-based intelligent recognition method for snow tracks and extreme environments, characterized in that, It includes the following steps: S1. A multi-modal data acquisition module, including a high-definition camera, an infrared camera, and a lidar, is used to synchronously acquire visible light images, infrared thermal images, and three-dimensional point cloud data of the snow environment; S2. A fusion-based deep learning model extracts spatial features by a convolutional neural network (CNN), models temporal dynamics by a recurrent neural network (RNN), and fuses multi-source data features through a cross-modal attention mechanism; S3. An adversarial generation data augmentation unit uses a conditional generative adversarial network (cGAN) to generate synthetic data of extreme snow scenarios, including virtual trajectory samples under blizzards, reflective interference, and low contrast conditions; S4. An adaptive control subsystem integrates an inertial measurement unit (IMU) and a road surface adhesion coefficient estimator to dynamically adjust the driving speed, suspension height, and sensor sampling frequency of the mobile platform; S5. An edge computing node deploys a lightweight model inference engine to achieve low-latency trajectory feature extraction and classification decision-making.

2. The intelligent recognition method for snow tracks and extreme environments based on AI according to claim 1, wherein The multi-modal data acquisition module in S1 includes: a polarization imaging unit for suppressing snow mirror reflection interference; a multi-baseline stereo vision array with camera groups configured with different field-of-view angles to achieve wide-range coverage; and an automatic calibration device for the lidar and visible light image to achieve sub-pixel level spatial alignment through a checkerboard target.

3. The intelligent recognition method for snow tracks and extreme environments based on AI according to claim 1, wherein, The fusion-based deep learning model in S2 adopts a multi-scale feature pyramid structure, including: parallel residual CNN branches for processing different resolution image inputs; a bidirectional gated recurrent unit (Bi-GRU) for capturing trajectory continuity features; The feature fusion layer realizes cross-modal feature alignment through learnable weights, and its weight coefficients are dynamically adjusted by environmental visibility parameters.

4. An AI-based intelligent recognition method for snow tracks and extreme environments according to claim 1, characterized in that, The training method of the fusion-based deep learning model in S2 is as follows: S2.

1. Adopt a curriculum learning strategy to train the model in stages according to the severity of the environment; S2.

2. Introduce a contrast loss function to enhance the tightness of similar trajectory features; S2.

3. Use differentiable data augmentation technology to achieve end-to-end robustness optimization.

5. The AI-based intelligent recognition method for snow tracks and extreme environments according to claim 1, characterized in that, The adversarial generation data augmentation unit in S3 includes: a physical constraint generator that integrates a fluid mechanics model to simulate the movement trajectory of real snow grains; the discriminator network adopts a multi-scale PatchGAN architecture to evaluate both the global consistency and local texture authenticity of the image; and a data augmentation strategy controller that automatically adjusts the difficulty level of the generated samples according to the model training error distribution.

6. The intelligent recognition method for snow tracks and extreme environments based on AI according to claim 1, wherein The adaptive control subsystem in S4 includes: a vibration suppression algorithm based on road surface roughness spectrum estimation to adjust the stability of the pan-tilt through feedforward-feedback composite control; a dynamic exposure control module that optimizes the shutter speed and ISO parameters of the camera in real time according to the snow reflectivity; and a lidar scan mode switching mechanism that automatically enables a high-density scan mode when the visibility is lower than the threshold.

7. An AI-based intelligent recognition method for snow tracks and extreme environments according to claim 1, characterized in that The edge computing node in S5 adopts: model distillation technology to transfer the decision-making knowledge of the teacher network to the lightweight student network; a dynamic computing offloading mechanism to allocate computing tasks among the edge, cloud, and device according to the network bandwidth; and a data caching strategy based on spatio-temporal correlation to preferentially retain the original data of abnormal trajectory segments.

8. The intelligent recognition method for snow tracks and extreme environments based on AI according to claim 1, wherein The adaptive control subsystem integrates an abnormal trajectory early warning module, and its implementation method is as follows: S4.

1. Construct a trajectory dynamics feature template library, including typical abnormal modes such as vehicle skidding and pedestrian falling; S4.

2. Design a spatio-temporal graph convolutional network (ST-GCN) to detect abnormal trajectory continuity; S4.

3. Generate a comprehensive early warning level by fusing multi-sensor confidence scores.

9. A system for an AI-based intelligent recognition method of snow tracks and extreme environments according to any one of claims 1-8, characterized in that, The application method of the system on a mobile platform includes: Step 1. Establish a snow adhesion coefficient - safe speed mapping table, and dynamically optimize path planning through reinforcement learning; Step 2. Deploy a vision-inertial tightly coupled SLAM system to achieve precise positioning in a GPS-free environment; Step 3. Design a modular thermal management device to ensure the working stability of each sensor under extreme low temperatures.

Images