YOLOv11 algorithm model for detecting and identifying pedestrians, vehicles and traffic signs in real time in automatic driving vehicle

Abstract

The invention provides a YOLOv11 algorithm model for detecting and identifying pedestrians, vehicles and traffic signs in real time in an automatic driving vehicle. The YOLOv11 algorithm model comprises a data acquisition module, a data preprocessing module, a feature extraction and object detection module, a reasoning optimization module and an output interface module, wherein the feature extraction and object detection module is a deep convolutional neural network module, and performs feature extraction and object positioning based on a YOLOv11 architecture; the deep convolutional neural network module can better adjust the position of a prediction frame by introducing an improved IoU loss function so as to more accurately position various target objects.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles and traffic signs in autonomous vehicles. Background Technology

[0002] Autonomous driving technology has developed rapidly in recent years. Its core lies in the vehicle's ability to autonomously perceive its surroundings and make decisions. Image recognition technology is a crucial component of the visual perception in autonomous driving systems, enabling the effective detection and recognition of roads, pedestrians, vehicles, and various traffic signs. Traditional image recognition methods rely on manual feature extraction, which struggles to achieve the high accuracy and real-time performance required in complex and ever-changing traffic environments. Current deep learning methods, such as Convolutional Neural Networks (CNNs), have achieved significant results in computer vision. These networks utilize convolutional and pooling layers to efficiently extract image features and identify targets. The YOLOv11 algorithm employs a single-stage detection principle, directly transforming the input image into detection results. This process involves segmenting the image into a grid and performing bounding box and class prediction on each grid. Mathematically, convolution operations can be represented by the formula: ; Where * denotes convolution operation, τ is the integration variable, and t is the independent variable. f and g are two functions that need to be convolved, and the integration process reflects the interaction between these two functions at a certain moment t.

[0003] The YOLO (You Only Look Once) model, as a representative of this approach, has garnered widespread attention for its excellent performance in object detection. It significantly improves processing speed by performing feature prediction and object localization in a single stage. However, with the increasing complexity of traffic environments, existing versions of YOLO still face challenges in balancing accuracy and real-time performance. In autonomous driving applications, image recognition tasks demand extremely high detection speed and accuracy, while existing technologies often fall short of accuracy while meeting real-time requirements. In some edge cases, the recognition of pedestrians or traffic signs may be inaccurate. Simultaneously, the complex and diverse traffic environments place higher demands on the model's generalization ability, and traditional techniques struggle to balance recognition accuracy and system response speed. In the field of object detection, the YOLO series of models is widely popular for its real-time performance and efficiency. However, existing YOLO models have several significant limitations in their loss function design, resulting in restricted detection performance. First, the traditional YOLO loss function relies on mean squared error (MSE) for bounding box localization, which often fails to accurately reflect the actual error in cases of severe occlusion or non-overlapping boxes, thus affecting localization accuracy. Secondly, classification loss typically uses cross-entropy, which can lead to poor detection of minor classes when dealing with imbalanced datasets. Furthermore, the design of confidence loss fails to adequately address background noise and the need for adaptive model adjustments. These shortcomings highlight the need for an improved loss function design to enhance overall detection accuracy and robustness.

[0004] Therefore, an improved target detection method is needed to achieve efficient and accurate real-time detection and recognition in complex traffic scenarios. This invention proposes a solution based on the YOLOv11 algorithm, aiming to enhance the visual perception capabilities of autonomous vehicles. By optimizing model design and inference processes, it overcomes the limitations of existing technologies and provides more reliable and faster detection results. Summary of the Invention

[0005] The purpose of this invention is to propose a YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles and traffic signs in autonomous vehicles, so as to overcome the limitations of existing technologies and provide more reliable and faster detection results.

[0006] To achieve the above objectives, this invention proposes a YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles, and traffic signs in autonomous vehicles. The model is characterized by comprising a data acquisition module, a data preprocessing module, a feature extraction and object detection module, an inference optimization module, and an output interface module. The data acquisition module is used to acquire image data from the environment surrounding the autonomous vehicle; The data preprocessing module is used to adjust the input image to meet the input requirements of the model; The feature extraction and object detection module is a deep convolutional neural network module, which performs feature extraction and object localization based on the YOLOv11 architecture; The inference optimization module improves inference speed to meet the real-time requirements of autonomous driving; The output interface module is used to output the recognition results to the autonomous vehicle control system to enable environmental perception and decision support for the vehicle. The deep convolutional neural network module, by introducing an improved intersection-over-union (IoU) loss function, is able to better adjust the position of the prediction box, thereby more accurately locating various target objects.

[0007] Furthermore, the real-time inference engine optimization model employs hardware acceleration technology or inference engine optimization libraries to improve computational efficiency.

[0008] Furthermore, the data preprocessing module includes image scaling, cropping, rotation, and color adjustment to improve the model's generalization ability.

[0009] Furthermore, the improved Intersection over Union (IoU) loss function is calculated as follows: ; ; Similar to IoU, GIoU is also a distance metric. GIoU ranges from -1 to 1, while GIoU Loss ranges from 0 to 2. GIoU is the lower bound of IoU; when two bounding boxes overlap infinitely, IoU = GIoU = 1. Unlike IoU, which only focuses on overlapping regions, GIoU considers not only overlapping regions but also other non-overlapping regions.

[0010] Compared with the prior art, the advantages of the present invention are: 1. The feature extraction and object detection module includes a convolutional neural network-based architecture. Optimizing this network architecture can better identify objects and improve the speed and accuracy of detection.

[0011] 2. The inference optimization module reduces data conversion and increases memory bandwidth and access speed by using hardware acceleration technology or inference engine optimization libraries. This enables more efficient use of computing resources and optimization of algorithm execution, further improving computational efficiency.

[0012] 3. The data preprocessing module performs data augmentation operations, including but not limited to image scaling, cropping, rotation, and color adjustment, which increases data diversity, can simulate different environmental conditions, and thus adapt to different environments and objectives, thereby improving the model's generalization ability.

[0013] 4. Convolutional neural network models use transfer learning strategies to fine-tune a pre-trained YOLO model. Since the pre-trained model already has good feature extraction capabilities, fine-tuning only requires adjusting a small part of the model, which can significantly reduce training time and computational resource consumption, thereby improving detection performance in different scenarios.

[0014] 5. This invention optimizes the loss function in the YOLO model to improve the accuracy and robustness of object detection. By introducing an improved Intersection over Union (IoU) loss function, the model can better adjust the position of the predicted bounding box, thereby more accurately locating various target objects. Attached Figure Description

[0015] Figure 1 This is a system framework diagram in an embodiment of the present invention; Figure 2 This is a schematic diagram of vehicle data collection in an embodiment of the present invention; Figure 3 Here is a diagram of the YOLOv11 model structure; Figure 4 Data processing flowchart: Figure 5 This is a schematic diagram of the inference optimization module. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0017] like Figure 1 As shown, this invention proposes a YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles, and traffic signs in autonomous vehicles. The workflow and structure are as follows: Data acquisition module, such as Figure 2 The diagram illustrates the layout of cameras and sensors on a vehicle; these cameras and sensors are used to acquire image data from the environment surrounding the autonomous vehicle. The data preprocessing module is connected to the data acquisition module, such as... Figure 4 As shown, it is used to process the acquired image data into a format suitable for deep learning models, including but not limited to image scaling, cropping, rotation, and color adjustment. It undertakes the key data augmentation task to comprehensively improve the generalization ability of the deep learning model YOLOv11, enabling it to cope with the complex and ever-changing situations in autonomous driving scenarios.

[0018] The feature extraction and object detection module, based on the YOLOv11 algorithm, is configured to extract features from the preprocessed image data and detect target objects, including pedestrians, vehicles, and traffic signs. By constructing a deep, optimized convolutional neural network architecture suited to autonomous driving conditions, it improves detection speed and accuracy, providing timely and accurate target information for intelligent driving decisions and effectively safeguarding driving safety. A transfer learning strategy is used to fine-tune a pre-trained YOLO model to improve detection performance in different scenarios. The feature extraction and object detection module is a deep convolutional neural network module, based on the YOLOv11 architecture for feature extraction and object localization. The model structure of the YOLOv11 algorithm is as follows: Figure 3 As shown.

[0019] This model integrates an improved loss function, including GIoU for bounding box localization. During model training, the weights of each component of the loss function are dynamically adjusted according to different training stages to optimize overall performance. The formula for calculating the GIoU loss is: ; ; Similar to IoU, GIoU is also a distance metric. GIoU ranges from -1 to 1, while GIoU Loss ranges from 0 to 2. GIoU is the lower bound of IoU; when two bounding boxes overlap infinitely, IoU = GIoU = 1. Unlike IoU, which only focuses on overlapping regions, GIoU considers not only overlapping regions but also other non-overlapping regions.

[0020] This design effectively improves localization accuracy by penalizing non-overlapping regions. In practical implementation, the network's loss function integrates this IoU variant, and experiments have verified that it outperforms the traditional MSE in convergence speed and accuracy on different datasets.

[0021] Preferably, the inference optimization module is connected to the feature extraction and object detection modules to accelerate the inference process of the object detection model and ensure real-time performance. Furthermore, by using hardware acceleration technology or an inference engine optimization library, computational efficiency can be rapidly improved. Its specific workflow is as follows: Figure 5 As shown.

[0022] The output interface module is used to output the recognition results to the autonomous vehicle control system to enable environmental perception and decision support for the vehicle. Once the feature extraction and object detection modules accurately locate pedestrians, vehicles, and traffic signs, the output interface module immediately operates efficiently, outputting the recognition results to the autonomous vehicle control system with an ultra-low latency of less than 10 milliseconds, according to a preset and compatible communication protocol, such as the CAN bus protocol. Based on this real-time and accurate environmental perception data, the control system quickly plans the vehicle's driving path, precisely controls the vehicle speed, brakes promptly when pedestrians suddenly appear ahead, flexibly changes lanes when vehicles cut in, or smoothly turns according to traffic signs, providing comprehensive and reliable decision support for autonomous driving and greatly improving driving safety and smoothness.

[0023] In the autonomous driving detection system of this invention, the working principle involves the coordinated operation of multiple modules. First, the vehicle-mounted camera captures images of the road environment in real time and transmits the data to a preprocessing module for basic image processing and formatting. Then, the preprocessed image is input to the core YOLOv11 algorithm module, which extracts key features of the image through a deep neural network and identifies targets such as pedestrians, vehicles, and traffic signs. The identification results are processed by an inference optimization module to accelerate detection speed and improve accuracy. Finally, the detection results are transmitted to an output interface module, converted into signals that the vehicle control system can understand, thereby assisting the vehicle in making real-time decisions, such as adjusting speed or path, to ensure driving safety. These modules work together to achieve environmental perception and response.

[0024] Through this system architecture, the present invention can significantly improve the real-time detection capability of autonomous vehicles in complex traffic environments, solve the problem of balancing accuracy and speed in existing technologies, and provide more reliable environmental perception and decision support for contemporary autonomous driving systems.

[0025] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles, and traffic signs in autonomous vehicles, characterized in that, It includes a data acquisition module, a data preprocessing module, a feature extraction and object detection module, an inference optimization module, and an output interface module; The data acquisition module is used to acquire image data from the environment surrounding the autonomous vehicle. The data preprocessing module is used to adjust the input image to meet the input requirements of the model; The feature extraction and object detection module is a deep convolutional neural network module, which performs feature extraction and object localization based on the YOLOv11 architecture. The inference optimization module improves inference speed to meet the real-time requirements of autonomous driving; The output interface module is used to output the recognition results to the autonomous vehicle control system to realize environmental perception and decision support for the vehicle. The deep convolutional neural network module, by introducing an improved intersection-over-union (IoU) loss function, can better adjust the position of the prediction box, thereby more accurately locating various target objects.

2. The YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles, and traffic signs in autonomous vehicles according to claim 1, characterized in that, The real-time inference engine optimization model employs hardware acceleration technology or inference engine optimization libraries to improve computational efficiency.

3. The YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles, and traffic signs in autonomous vehicles according to claim 1, characterized in that, The data preprocessing module includes image scaling, cropping, rotation, and color adjustment to improve the model's generalization ability.

4. The YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles, and traffic signs in autonomous vehicles according to claim 1, characterized in that, The improved Intersection over Union (IoU) loss function is calculated as follows: ； ； Similar to IoU, GIoU is also a distance metric. GIoU ranges from [-1, 1], while GIoU Loss ranges from [0, 2]. GIoU is the lower bound of IoU; when two bounding boxes overlap infinitely, IoU = GIoU = 1. Unlike IoU, which only focuses on overlapping regions, GIoU considers not only overlapping regions but also other non-overlapping regions.

5. The YOLOv11 algorithm model for real-time detection and recognition of pedestrians, vehicles, and traffic signs in autonomous vehicles according to claim 1, characterized in that, The deep convolutional neural network module employs a transfer learning strategy.

Images