Auxiliary lane changing method and system based on semantic analysis of vehicle tail lamp

Abstract

The invention belongs to but is not limited to the technical field of intelligent driving, and particularly relates to an auxiliary lane changing method and system based on semantic analysis of vehicle taillights, and the method comprises the steps: S1, constructing a data set; s2, data annotation and classification; s3, model construction and optimization improvement; s4, optimizing a target detection head; and S5, constructing and optimizing a loss function.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the field of intelligent driving technology, and in particular relates to an auxiliary lane changing method and system based on semantic analysis of vehicle taillights. Background Art

[0002] With the rapid growth of the global car ownership, the complexity of road traffic scenes has also increased dramatically. In order to make safe driving decisions in road driving scenarios to complete lane changes, steering, overtaking and other behaviors, drivers need to obtain rich information, including driving semantic information.

[0003] Among these driving semantic information, vehicle taillights are one of the most commonly used and important features. Different combinations of taillight signals can convey the driving intention of the front vehicle to the rear vehicle. By observing the taillights of the front vehicle, the driver of the rear vehicle can obtain information such as the steering and lane change of the front vehicle.

[0004] For example, the vehicle's turn signal flashes to indicate that the vehicle is about to change lanes or turn, and the brake light turns on to indicate that the vehicle is slowing down or stopping. These taillight signals are important bases for drivers to make driving decisions.

[0005] Therefore, visual perception plays a vital role in driver assistance systems. Using visual perception technology to analyze vehicle taillight signals can improve the accuracy of detecting the lane change intention of the vehicle in front, thereby effectively assisting the driver to drive safely and reduce the occurrence of traffic accidents.

[0006] In this context, vision-based vehicle lane change detection technology is crucial in the field of intelligent driving.

[0007] Although some existing intelligent driving assistance systems can provide basic road information and auxiliary functions, they still have problems such as insufficient accuracy and low real-time performance in detecting vehicle lane change intentions. These systems often rely on simple visual features and cannot fully understand and analyze complex taillight signal combinations, resulting in unsatisfactory performance in complex traffic environments.

[0008] Therefore, it is urgent to design a vehicle taillight lane change detection system based on advanced visual perception technology, which can improve the accuracy and real-time detection of vehicle lane change intentions by analyzing the semantic information of taillight signals, and further improve driving safety. This will provide important support for the development of intelligent driving technology, and will also have a positive impact on reducing traffic accidents and improving traffic conditions.

[0009] In view of the above analysis, the technical problems that need to be solved urgently in the existing technology are: although some existing intelligent driving assistance systems can provide basic road information and auxiliary functions, they still have insufficient accuracy and low real-time performance in detecting vehicle lane change intentions. These systems often rely on simple visual features and cannot fully understand and analyze complex taillight signal combinations, resulting in unsatisfactory performance in complex traffic environments. Summary of the invention

[0010] In view of the problems existing in the prior art, the present invention provides an auxiliary lane changing method and system based on semantic analysis of vehicle taillights.

[0011] The present invention is implemented as follows: a lane-changing assistance method based on semantic analysis of vehicle taillights, comprising: S1: Dataset construction; S2: Data annotation and classification; S3: Model building and optimization improvements; S4: object detection head optimization; S5 Loss function construction and optimization.

[0012] Furthermore, S1 specifically includes: Existing dataset: Based on the BDD100K dataset, 2,000 images containing taillight features are selected; Field shooting: 2,000 images were taken on site at different road sections, times, and environments to build a vehicle taillight dataset; Data diversity: The dataset covers a variety of weather, time periods, and road conditions, meeting the generalization requirements of model training.

[0013] Furthermore, S2 specifically includes: Labeling tool: Use LabelImg software to label images; Classification labels: Assign category labels to each bounding box, including brake lights, left and right turn signal lights, and vehicles for taillights; Dataset division: Use Python script to divide the labeled dataset into training set (3200 images), validation set (600 images) and test set (200 images).

[0014] Furthermore, S3 specifically includes: Benchmark model: YOLOv8 is selected as the benchmark model, and optimized for the real-time, lightweight, and efficient requirements of the taillight detection task; Optimize sampling operators: Bilinear Interpolation with Depthwise Separable Convolutions: Upsampling using bilinear interpolation; Combining depth-wise separable convolution reduces the number of parameters and computational overhead, and improves the resolution and clarity of feature maps; Attention Mechanism: Squeeze-and-Excitation Networks (SENet): Through compression and expansion operations, feature channels are automatically learned and weighted to improve the representation of important features and suppress unimportant features; Feature fusion method: Feature Pyramid Networks (FPN): A feature pyramid network is used to enhance the fusion and utilization of features of different scales, improve the model's detection capability when dealing with taillight targets of different sizes, and improve the overall detection performance.

[0015] Furthermore, S4 specifically includes: Reconstructed detection head: Remove the channel compression step of the original detection head to improve the convolutional layer's ability to learn feature information; Introducing depthwise separable convolution: optimizing parameter operation logic, reducing network parameters, and improving computational efficiency; Ordinary convolution calculation: Calculation amount = n·k·k·m·p·p Depthwise separable convolution computation: Calculation amount = k·k·m·p·p + n·m·p·p Parameter calculation savings: The amount of computation saved = n·k·k·m·p·p-(k·k·m·p·p+n·m·p·p).

[0016] Furthermore, S5 specifically includes: Complete Intersection over Union (CIoU): The CIoU metric is introduced to comprehensively consider the scale and position relationship, more accurately describe the position and size characteristics of the taillights, and improve detection accuracy and robustness.

[0017] Furthermore, model training specifically includes: 1) Data enhancement and hybrid training: Data enhancement: During the training process, a variety of data enhancement techniques are applied, including random cropping, rotation, color transformation, lighting changes, etc., to enhance the model's adaptability to different environments and conditions; Hybrid training: Combining transfer learning and self-supervised learning methods, using a pre-trained model (such as YOLOv8 trained on a large dataset) as a starting point for refined training. Self-supervised learning can be used to further optimize the model's performance in the taillight detection task; Data augmentation: Use the `albumentations` library to augment training data during loading; Transfer learning: load the pre-trained YOLOv8 weights and fine-tune them on the taillight dataset; Self-supervised learning: Introducing self-supervised learning algorithms, such as contrastive learning, to improve the model’s ability to learn taillight features; 2) Advanced optimization strategies: Learning rate scheduling: Use learning rate scheduling strategies (such as cosine annealing or step annealing) to dynamically adjust the learning rate to avoid overfitting during training; Mixed precision training: Use mixed precision training (FP16) to reduce memory consumption and speed up training while maintaining model accuracy; Learning rate scheduling: implemented through PyTorch's `lr_scheduler` module; Mixed precision training: Use the torch.cuda.amp module.

[0018] Furthermore, model evaluation specifically includes: 1) Comprehensive evaluation indicators: Precision and recall: In addition to accuracy, the model's precision, recall, and F1 score are calculated to comprehensively evaluate the model's performance. Real-time performance evaluation: measuring the response time and inference speed of the model in actual applications to verify the effectiveness of real-time detection; Comprehensive evaluation: Use the evaluation indicators in the scikit-learn library to calculate the function; Real-time performance evaluation: testing the model’s inference speed and latency in an actual hardware environment; 2) Anomaly detection and robustness testing: Anomaly detection: Introduce anomaly detection mechanisms to identify and handle the performance of the model under abnormal conditions; Robustness testing: Test the model in different lighting, weather conditions, and scenarios to ensure its robustness; Anomaly detection: Use `Isolation Forest` or `One-Class SVM` for anomaly detection; Robustness testing: Performance verification in a variety of test environments.

[0019] Furthermore, model deployment specifically includes: 1) Optimize model conversion: ONNX optimization: Use ONNX optimization tools (such as ONNX Runtime) to optimize the model and reduce the computational overhead during inference; TensorRT optimization: Apply TensorRT's dynamic quantization and layer fusion technology to further accelerate model reasoning; ONNX optimization: use the `onnxruntime` library of ONNX Runtime; TensorRT optimization: Apply TensorRT's `trt` library for optimization; 2) Edge computing and cloud deployment: Edge computing: Deploy models on AI edge devices and use edge computing capabilities to achieve low-latency real-time detection; Cloud backup: Deploy models to cloud servers to support centralized processing and large-scale data analysis; Edge computing: using edge devices such as NVIDIA Jetson series or Google Coral; Cloud backup: Use AWS Sagemaker or Google Cloud AI Platform for model management and backup.

[0020] Another object of the present invention is to provide an assisted lane changing system based on vehicle taillight semantic analysis for realizing the assisted lane changing method based on vehicle taillight semantic analysis, comprising: Dataset construction module, used to construct the dataset; Data labeling and classification module, used to label data and classify the labeled data; Model building module, used for model building and optimization improvement; Target detection head optimization module, used for optimizing the target detection head; Loss function optimization module, used to construct and optimize the loss function.

[0021] Another object of the present invention is to provide a computer device, the computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the assisted lane change method based on vehicle taillight semantic analysis.

[0022] Another object of the present invention is to provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the assisted lane change method based on vehicle taillight semantic analysis.

[0023] Another object of the present invention is to provide an information data processing terminal, which includes the auxiliary lane change system based on vehicle taillight semantic analysis.

[0024] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows: First, the technical problems of the prior art solved by the technical solution of the present invention are: 1) Insufficient accuracy of taillight signal detection: Existing technical problems: In complex driving environments, such as changing weather, lighting conditions and different road conditions, the detection accuracy of vehicle taillight signals is often insufficient. This may cause the autonomous driving or driving assistance system to be unable to accurately identify the brake light and turn signal signals of the vehicle in front, thereby increasing the risk of accidents.

[0025] The solution of the present invention: by constructing a diverse taillight data set covering different weather, time periods and road conditions, and adopting an optimized YOLOv8 model and feature fusion method, the present invention significantly improves the accuracy of taillight signal detection and ensures that taillight signals can be reliably identified in various complex environments.

[0026] 2) The model’s generalization ability in different environments is insufficient: Existing technical problems: The generalization ability of traditional taillight detection models in different environments is poor, especially in low-light or strong light interference environments, the performance of the detection model decreases significantly. This limits the application of the model in various scenarios and reduces the practicality of the driver assistance system.

[0027] The solution of the present invention: The present invention improves the feature extraction and feature fusion capabilities of the model in different environments by introducing the SENet attention mechanism and the FPN feature pyramid network, enhances the generalization ability of the model, enables it to adapt to a variety of complex scenarios, and provides stable detection performance.

[0028] 3) The computational efficiency and real-time performance of the taillight detection model are insufficient: Existing technical problems: In practical applications, the taillight detection model needs to achieve real-time detection under limited computing resources. However, due to the large number of parameters and high computing overhead, the existing methods are difficult to meet the real-time requirements, especially on resource-constrained embedded devices.

[0029] The solution of the present invention: By optimizing the sampling operator (such as using bilinear interpolation combined with depthwise separable convolution) and reconstructing the detection head, the present invention significantly reduces the number of parameters and computational overhead of the model, improves the computational efficiency of the model, and enables taillight detection to run in real time on embedded devices.

[0030] The significant technical progress obtained by the present invention: 1) Significantly improved the accuracy of taillight detection: The present invention significantly improves the accuracy of taillight signal detection through diversified data set construction and model optimization, especially model training under different lighting and weather conditions. Whether it is daytime, nighttime, or in adverse weather conditions such as rainy days or haze, the model can reliably detect vehicle taillight signals, improving the safety of the driver assistance system.

[0031] 2) Enhanced the model’s generalization ability and environmental adaptability: By introducing the SENet attention mechanism and the FPN feature pyramid network, the present invention enhances the model's adaptability to different environments, allowing taillight detection to maintain high performance in a variety of scenarios. This advancement enables the model to be widely used in a variety of driving scenarios, thereby improving the practicality of autonomous driving or driver assistance systems.

[0032] 3) Improved the computational efficiency of the model and achieved real-time detection: The present invention significantly reduces the computational complexity and parameter quantity of the model through deep separable convolution and optimized detection head design, improves computational efficiency, and enables the model to achieve real-time detection on resource-constrained embedded devices. This progress is crucial for the real-time requirements in autonomous driving systems and ensures the system's rapid response capability.

[0033] 4) Comprehensively optimize the overall performance of taillight detection: The present invention has comprehensively optimized the accuracy, generalization ability and computational efficiency of the model, which not only improves the overall performance of taillight detection, but also ensures its stable performance in a variety of complex driving environments. This comprehensive technological advancement provides solid technical support for the safety and reliability of autonomous driving systems and driver assistance systems.

[0034] Through the technical solution of the present invention, the problems of insufficient accuracy, generalization ability and real-time performance of taillight detection in the prior art have been effectively solved. The significant technological progress enables the taillight detection system to work stably in complex environments, ensuring the safety and reliability of autonomous driving and driver assistance systems, thereby promoting the development of intelligent driving technology.

[0035] Second, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects: (1) The technical solution of the present invention solves a technical problem that people have been eager to solve but have never been able to solve successfully: Through the technical solution of the present invention, the problems of insufficient accuracy, generalization ability and real-time performance of taillight detection in the prior art have been effectively solved. The significant technological progress enables the taillight detection system to work stably in complex environments, ensuring the safety and reliability of autonomous driving and driver assistance systems, thereby promoting the development of intelligent driving technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of an auxiliary lane changing method based on vehicle taillight semantic analysis provided by an embodiment of the present invention; Figure 2 It is a structural diagram of an auxiliary lane change system based on vehicle taillight semantic analysis provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0038] like Figure 1 As shown, the lane-changing assistance method based on vehicle taillight semantic analysis provided by an embodiment of the present invention includes: S1: Dataset Construction Existing dataset: Based on the BDD100K dataset, 2,000 images containing rich taillight features are selected.

[0039] Field shooting: 2,000 images were taken on site at different road sections, times, and environments to build a diverse vehicle taillight dataset.

[0040] Data diversity: The dataset covers a variety of weather, time periods, and road conditions, meeting the generalization requirements of model training.

[0041] S2: Data labeling and classification Labeling tool: Use LabelImg software to label images.

[0042] Classification labels: Assign category labels to each bounding box, including brake lights, left and right turn signal lights, and vehicles for taillights.

[0043] Dataset division: Use Python script to divide the labeled dataset into training set (3200 images), validation set (600 images) and test set (200 images).

[0044] S3: Model building and optimization improvements Benchmark model: YOLOv8 is selected as the benchmark model and optimized for the real-time, lightweight, and efficient requirements of the taillight detection task.

[0045] Optimize sampling operators: Bilinear Interpolation with Depthwise Separable Convolutions: Upsampling is done using bilinear interpolation.

[0046] Combining depth-wise separable convolutions reduces the number of parameters and computational overhead, and improves the resolution and clarity of feature maps.

[0047] Attention Mechanism: Squeeze-and-Excitation Networks (SENet): Through compression and expansion operations, feature channels are automatically learned and weighted to improve the representation ability of important features and suppress unimportant features.

[0048] Feature fusion method: Feature Pyramid Networks (FPN): A feature pyramid network is used to enhance the fusion and utilization of features of different scales, improve the model's detection capability when dealing with taillight targets of different sizes, and improve the overall detection performance.

[0049] S4: Object detection head optimization Reconstructed detection head: Remove the channel compression step of the original detection head to improve the convolutional layer's ability to learn feature information.

[0050] Introducing depthwise separable convolution: optimizing parameter operation logic, reducing network parameters, and improving computational efficiency.

[0051] Ordinary convolution calculation: Calculation amount = n·k·k·m·p·p Depthwise separable convolution computation: Calculation amount = k·k·m·p·p + n·m·p·p Parameter calculation savings: The amount of computation saved = n·k·k·m·p·p-(k·k·m·p·p+n·m·p·p).

[0052] S5 Loss function construction and optimization Complete Intersection over Union (CIoU): The CIoU metric is introduced to comprehensively consider the scale and position relationship, more accurately describe the position and size characteristics of the taillights, and improve detection accuracy and robustness; CloU calculation formula:

[0053] in, represents the distance from the center point, represents the diagonal distance, and A measure of aspect ratio differences.

[0054] Model training specifically includes: 1) Data enhancement and hybrid training Data augmentation: During the training process, a variety of data augmentation techniques are applied, such as random cropping, rotation, color transformation, lighting changes, etc., to enhance the model's ability to adapt to different environments and conditions.

[0055] Hybrid training: Combining transfer learning and self-supervised learning methods, using a pre-trained model (such as YOLOv8 trained on a large dataset) as a starting point for refinement training. Self-supervised learning can be used to further optimize the model's performance in the taillight detection task.

[0056] Data augmentation: Use the `albumentations` library to augment the training data during loading.

[0057] Transfer learning: load pre-trained YOLOv8 weights and fine-tune on the taillight dataset.

[0058] Self-supervised learning: Introduce self-supervised learning algorithms, such as contrastive learning, to improve the model's ability to learn taillight features.

[0059] 2) Advanced optimization strategies Learning rate scheduling: Use learning rate scheduling strategies (such as cosine annealing or step annealing) to dynamically adjust the learning rate to avoid overfitting during training.

[0060] Mixed Precision Training: Use mixed precision training (FP16) to reduce memory consumption and speed up training while maintaining model accuracy.

[0061] Learning rate scheduling: implemented through PyTorch's `lr_scheduler` module.

[0062] Mixed precision training: Use the torch.cuda.amp module.

[0063] Model evaluation specifically includes: 1) Comprehensive evaluation indicators Precision and recall: In addition to accuracy, the model's precision, recall, and F1 score are calculated to comprehensively evaluate the model's performance.

[0064] Real-time performance evaluation: Measuring the response time and inference speed of the model in real applications to verify the effectiveness of real-time detection.

[0065] Comprehensive evaluation: functions are evaluated using evaluation metrics from the scikit-learn library.

[0066] Real-time performance evaluation: Test the inference speed and latency of the model in an actual hardware environment.

[0067] 2) Anomaly detection and robustness testing Anomaly detection: Introduce anomaly detection mechanism to identify and handle the performance of the model under abnormal conditions.

[0068] Robustness testing: Test the model in different lighting, weather conditions, and scenarios to ensure its robustness.

[0069] Anomaly Detection: Use `Isolation Forest` or `One-Class SVM` for anomaly detection.

[0070] Robustness testing: Performance verification in a variety of test environments.

[0071] Model deployment specifically includes: 1) Optimize model conversion ONNX optimization: Use ONNX optimization tools (such as ONNX Runtime) to optimize the model and reduce the computational overhead during inference.

[0072] TensorRT optimization: Apply TensorRT's dynamic quantization and layer fusion technology to further accelerate model reasoning.

[0073] ONNX Optimization: Use the `onnxruntime` library of ONNX Runtime.

[0074] TensorRT optimization: Apply TensorRT's `trt` library for optimization.

[0075] 2) Edge computing and cloud deployment Edge computing: Deploy models on AI edge devices and use edge computing capabilities to achieve low-latency real-time detection.

[0076] Cloud backup: Deploy models to cloud servers to support centralized processing and large-scale data analysis.

[0077] Edge computing: Use edge devices such as the NVIDIA Jetson series or Google Coral.

[0078] Cloud backup: Use AWS Sagemaker or Google Cloud AI Platform for model management and backup.

[0079] like Figure 2 As shown, the lane change assistance system based on vehicle taillight semantic analysis provided by the embodiment of the present invention includes: Dataset construction module, used to construct the dataset; Data labeling and classification module, used to label data and classify the labeled data; Model building module, used for model building and optimization improvement; Target detection head optimization module, used for optimizing the target detection head; Loss function optimization module, used to construct and optimize the loss function.

[0080] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of an assisted lane change method based on semantic analysis of vehicle taillights.

[0081] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of an assisted lane change method based on semantic analysis of vehicle taillights.

[0082] An application embodiment of the present invention provides an information data processing terminal, which includes an auxiliary lane change system based on semantic analysis of vehicle taillights.

[0083] Example 1: Assisted lane change method based on sunny highway scenario On a sunny highway, due to the high speed of vehicles, it is particularly important to accurately recognize the taillight signal when the vehicle changes lanes. This embodiment aims to assist the driver to change lanes safely on the highway through the semantic analysis of the vehicle taillights.

[0084] 1) Dataset construction: Existing dataset: Based on the BDD100K dataset, 1,000 vehicle taillight images on sunny highway scenes were selected to ensure that the dataset contains rich taillight features.

[0085] Field shooting: 1,000 vehicle taillight images were taken on different highway sections under sunny conditions to ensure that the dataset covers different vehicle models and headlight conditions.

[0086] Data diversity: The dataset contains images from different time periods (e.g., morning, afternoon, dusk), covering vehicle taillight features under various lighting conditions.

[0087] 2) Data labeling and classification: Labeling tools: All images were labeled using LabelImg software, focusing on brake and turn signal lights.

[0088] Classification label: Assign the labels of brake light, left turn signal, right turn signal and vehicle to each annotated tail light.

[0089] Dataset division: Use Python script to divide the labeled dataset into training set (1600 images), validation set (300 images) and test set (100 images).

[0090] 3) Model construction and optimization: Baseline model: YOLOv8 is selected as the detection model. The model is optimized through bilinear interpolation and depthwise separable convolution to improve the model's ability to detect taillights in highway scenarios.

[0091] Attention mechanism: The SENet mechanism is introduced to improve the model's attention to important features, suppress unimportant features, and optimize the detection accuracy of taillight signals through compression and expansion operations.

[0092] Feature fusion: Use the FPN feature pyramid network to enhance the fusion of features of different scales and improve the detection performance of taillight targets of different sizes.

[0093] 4) Target detection head optimization: Reconstruct detection head: optimize the detection head, remove the channel compression step of the original detection head, and improve feature learning capabilities.

[0094] Introducing depthwise separable convolution: Optimizing the convolution layer of the detection head and improving detection efficiency by reducing parameters.

[0095] 5) Loss function optimization: CIoU metric: The CIoU loss function is introduced to further improve the accuracy and robustness of detection by combining the position and size features of the taillights.

[0096] After training, the model's taillight detection accuracy in sunny highway scenarios reached over 95%, significantly improving the vehicle's ability to recognize the taillight signals of surrounding vehicles when changing lanes on the highway, assisting the driver in changing lanes safely.

[0097] Example 2: Assisted lane change method based on nighttime urban road scene The lighting conditions on urban roads at night are poor, which can easily interfere with the detection of vehicle taillights. This embodiment aims to improve the taillight signal recognition capability in urban road scenes at night by optimizing the taillight detection model, thereby assisting safe lane change.

[0098] 1) Dataset construction: Existing datasets: In the BDD100K dataset, 1,000 vehicle taillight images in night city road scenes were selected.

[0099] Field shooting: 1,000 vehicle taillight images were taken under different urban road and night lighting conditions to ensure that the taillight features captured from different angles and distances were covered.

[0100] Data diversity: The dataset includes different night lighting conditions, such as street lighting, high beam, low beam, and shadow areas, to ensure the generalization ability of model training.

[0101] 2) Data labeling and classification: Labeling tool: Use LabelImg to label all images, with an emphasis on the brake lights and turn signal lights of the taillights.

[0102] Classification label: Assign a corresponding category label to each taillight annotation, including brake light, left turn signal, right turn signal, and vehicle.

[0103] Dataset division: Use Python script to divide the dataset into training set (1600 images), validation set (300 images) and test set (100 images).

[0104] 3) Model construction and optimization: Baseline model: YOLOv8 is used as the baseline model. The model is optimized through bilinear interpolation and depthwise separable convolution to improve the model's adaptability to night scenes.

[0105] Attention mechanism: The SENet attention mechanism is introduced to enhance the model's attention to important features in night scenes and reduce the impact of lighting changes on taillight detection.

[0106] Feature fusion: The FPN feature pyramid network is applied to improve the model's ability to detect taillight targets at different distances and lighting conditions.

[0107] 4) Target detection head optimization: Reconstructed detection head: Optimize the detection head, remove the channel compression step of the original detection head, and improve the learning ability of taillight signals in low-light scenes at night.

[0108] Introducing depthwise separable convolution: Optimize the convolution layer of the detection head, reduce parameters, and improve computational efficiency.

[0109] 5) Loss function optimization: CIoU metric: The CIoU loss function is used to improve the accuracy and stability of detection by combining the position and size characteristics of tail light signals at night.

[0110] After training, the model's taillight detection accuracy in night-time urban road scenarios reached over 93%, significantly improving the ability to recognize vehicle taillight signals in low-light conditions and ensuring the safety of changing lanes at night.

[0111] 2. Relevant evidence of the technical effects obtained by the embodiments of the present invention.

[0112] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0113] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A lane-changing assistance method based on semantic analysis of vehicle taillights, characterized in that: include: S1: Dataset construction; S2: Data annotation and classification; S3: Model building and optimization improvements; S4: object detection head optimization; S5: Loss function construction and optimization.

2. The lane-changing assistance method based on vehicle taillight semantic analysis as claimed in claim 1, characterized in that: S1 specifically includes: Existing dataset: Based on the BDD100K dataset, 2,000 images containing taillight features are selected; Field shooting: 2,000 images were taken on site at different road sections, times, and environments to build a vehicle taillight dataset; Data diversity: The dataset covers a variety of weather, time periods, and road conditions, meeting the generalization requirements of model training.

3. The lane-changing assistance method based on vehicle taillight semantic analysis as claimed in claim 1, characterized in that: S2 specifically includes: Labeling tool: Use LabelImg software to label images; Classification labels: Assign category labels to each bounding box, including brake lights, left and right turn signal lights, and vehicles. Dataset division: Use Python script to divide the labeled dataset into training set, validation set and test set.

4. The lane-changing assistance method based on vehicle taillight semantic analysis as claimed in claim 1, characterized in that: S3 specifically includes: Benchmark model: YOLOv8 is selected as the benchmark model, and optimized for the real-time, lightweight, and efficient requirements of the taillight detection task; Optimize sampling operator: use bilinear interpolation for upsampling; Combining depth-wise separable convolution reduces the number of parameters and computational overhead, and improves the resolution and clarity of feature maps; Attention mechanism: Through compression and expansion operations, it automatically learns and weights feature channels, improves the representation ability of important features, and suppresses unimportant features; Feature fusion method: Use feature pyramid network to enhance the fusion and utilization of features of different scales, improve the model's detection ability when dealing with taillight targets of different sizes, and improve the overall detection performance.

5. The lane-changing assistance method based on vehicle taillight semantic analysis as claimed in claim 1, characterized in that: S4 specifically includes: Reconstructed detection head: Remove the channel compression step of the original detection head to improve the convolutional layer's ability to learn feature information; Introducing depthwise separable convolution: optimizing parameter operation logic, reducing network parameters, and improving computational efficiency; Ordinary convolution calculation: Calculation amount = n·k·k·m·p·p Depthwise separable convolution computation: Calculation amount = k·k·m·p·p + n·m·p·p Parameter calculation savings: The amount of computation saved = n·k·k·m·p·p-(k·k·m·p·p+n·m·p·p).

6. The lane-changing assistance method based on vehicle taillight semantic analysis as claimed in claim 1, characterized in that: S5 specifically includes: Complete intersection over union: Using the CIoU metric, it comprehensively considers the scale and position relationship to describe the position and size characteristics of the taillights.

7. The lane-changing assistance method based on vehicle taillight semantic analysis as claimed in claim 1, characterized in that: Model training specifically includes: 1) Data enhancement and hybrid training: Data enhancement: During the training process, a variety of data enhancement techniques are applied, including random cropping, rotation, color transformation, and lighting changes, to enhance the model's adaptability to different environments and conditions; Hybrid training: Combining transfer learning and self-supervised learning methods, using the pre-trained model as a starting point for refined training, and further optimizing the model's performance in the taillight detection task through self-supervised learning; Data augmentation: Use the `albumentations` library to augment training data during loading; Transfer learning: load the pre-trained YOLOv8 weights and fine-tune them on the taillight dataset; Self-supervised learning: Introducing self-supervised learning algorithms, including contrastive learning, to improve the model’s ability to learn taillight features; 2) Advanced optimization strategies: Learning rate scheduling: Use the learning rate scheduling strategy to dynamically adjust the learning rate to avoid overfitting during training; Mixed precision training: Use mixed precision training to reduce memory consumption and speed up training while maintaining model accuracy; Learning rate scheduling: implemented through PyTorch's `lr_scheduler` module; Mixed precision training: Use the torch.cuda.amp module.

8. The lane-changing assistance method based on vehicle taillight semantic analysis as claimed in claim 1, characterized in that: Model evaluation specifically includes: 1) Comprehensive evaluation indicators: Precision and recall: In addition to accuracy, the precision, recall, and F1 score of the model are calculated to comprehensively evaluate the model performance; Real-time performance evaluation: measuring the response time and inference speed of the model in actual applications to verify the effectiveness of real-time detection; Comprehensive evaluation: Use the evaluation indicators in the scikit-learn library to calculate the function; Real-time performance evaluation: testing the model’s inference speed and latency in an actual hardware environment; 2) Anomaly detection and robustness testing: Anomaly detection: Introduce anomaly detection mechanisms to identify and handle the performance of the model under abnormal conditions; Robustness testing: Test the model in different lighting, weather conditions, and scenarios to ensure its robustness; Anomaly detection: Use `Isolation Forest` or `One-Class SVM` for anomaly detection; Robustness testing: Performance verification in a variety of test environments.

9. The lane-changing assistance method based on vehicle taillight semantic analysis as claimed in claim 1, characterized in that: Model deployment specifically includes: 1) Optimize model conversion: ONNX optimization: Use ONNX optimization tools to optimize the model and reduce the computational overhead during inference; TensorRT optimization: Apply TensorRT's dynamic quantization and layer fusion technology to further accelerate model reasoning; ONNX optimization: use the `onnxruntime` library of ONNX Runtime; TensorRT optimization: Apply TensorRT's `trt` library for optimization; 2) Edge computing and cloud deployment: Edge computing: Deploy models on AI edge devices and use edge computing capabilities to achieve low-latency real-time detection; Cloud backup: Deploy models to cloud servers to support centralized processing and large-scale data analysis; Edge computing: using edge devices such as NVIDIA Jetson series or Google Coral; Cloud backup: Use AWS Sagemaker or Google Cloud AI Platform for model management and backup.

10. An assisted lane changing system based on vehicle taillight semantic analysis that implements the assisted lane changing method based on vehicle taillight semantic analysis as claimed in any one of claims 1 to 9, characterized in that: include: Dataset construction module, used to construct the dataset; Data labeling and classification module, used to label data and classify the labeled data; Model building module, used for model building and optimization improvement; Target detection head optimization module, used for optimizing the target detection head; Loss function optimization module, used to construct and optimize the loss function.

Images