Haze environment vehicle target detection system and method based on improved YOLO11

Abstract

The invention relates to the technical field of vehicle target detection, in particular to a haze environment vehicle target detection system and method based on improved YOLO11, and the method comprises the following steps: carrying out the defogging enhancement preprocessing of a haze image, and generating a clear image through an improved AOD-Net and Zero-DCE combined network; inputting the preprocessed image into an improved YOLO11 network, and enhancing vehicle feature extraction by using an SE attention mechanism; a self-adaptive multi-scale feature fusion strategy is adopted, and vehicle features of different scales are integrated; and performing target frame regression based on the optimized anchor frame and CIoU loss, and outputting a detection result. According to the invention, through defogging enhancement and a double-attention mechanism, vehicle features in haze are effectively extracted, and the omission ratio and the false drop rate are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle target detection, and in particular to a vehicle target detection system and method in a hazy environment based on an improved YOLO11. Background Art

[0002] With the rapid development of intelligent transportation systems (ITS), vehicle target detection, as a core technology, plays a vital role in numerous fields, including autonomous driving, intelligent monitoring, and traffic flow statistics. Accurate and efficient detection of vehicle targets is a key prerequisite for achieving intelligent traffic management and improving traffic safety. Under normal weather conditions, many established vehicle detection algorithms can achieve high accuracy and recall rates. However, when faced with complex natural environments, particularly foggy and hazy weather, the performance of these algorithms often degrades significantly. Hazy weather exacerbates light scattering and absorption, severely degrading image quality and blurring key features such as vehicle textures and edges, greatly increasing the difficulty of vehicle detection. Statistics show that the incidence of traffic accidents in foggy and hazy weather increases significantly compared to normal weather conditions, highlighting the urgency and importance of accurately detecting vehicles in adverse weather conditions. Currently, mainstream object detection algorithms, such as the region proposal-based R-CNN series and the single-stage YOLO series, face severe challenges when detecting vehicles in foggy and smoggy conditions, including image degradation and loss of target details. Traditional algorithms struggle to effectively extract features of vehicles obscured by smog, resulting in detection accuracy and speed that fall short of practical application requirements. YOLO11, a leading object detection algorithm, demonstrates excellent performance in a wide range of scenarios with its exceptional detection speed and high accuracy. It utilizes a more efficient network structure and optimization strategies, enabling rapid processing of large amounts of image data. However, the original YOLO11 algorithm was not designed to fully account for the impact of adverse weather conditions such as haze on images. Consequently, its performance improvement is limited when directly applied to vehicle detection in hazy conditions. Summary of the Invention

[0003] The purpose of the present invention is to provide a vehicle target detection system and method in a hazy environment based on an improved YOLO11 to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions: A vehicle target detection method in a hazy environment based on an improved YOLO11 includes the following steps: Perform dehazing and enhancement preprocessing on haze images, and generate clear images by improving the network combining AOD-Net and Zero-DCE; The preprocessed image is input into the improved YOLO11 network, and the SE attention mechanism is used to enhance vehicle feature extraction; Adopting an adaptive multi-scale feature fusion strategy to integrate vehicle features of different scales; Perform target box regression based on optimized anchor box and CIoU loss, and output detection results.

[0005] Preferably, the input haze image , by improving the network combining AOD-Net and Zero-DCE to generate dehazed enhanced images The specific process is as follows: Estimating fusion parameters using improved AOD-Net , through the formula Generate a preliminary defogging image, where b is a constant deviation with a value of 1. The calculation formula is: in, is the atmospheric projection rate, A is the atmospheric illumination value; The Zero-DCE module is introduced for low-light enhancement, and the DCE-Net output curve parameters are estimated through the depth curve. , adjust the image dynamic range pixel by pixel, and the enhanced image expression is: in, is the enhancement function based on the curve parameters, and n is the number of iterations.

[0006] Preferably, the defogging enhancement preprocessing includes: Introducing skip connection and SE-PA parallel channel attention mechanism in AOD-Net; Use the Zero-DCE module for low-light enhancement and adjust the image dynamic range through the depth curve estimation network; A composite loss function including SSIM, brightness and contrast is used to optimize the dehazing process.

[0007] Preferably, the feature fusion module of the improved YOLO11 network includes: Perform channel attention weighted fusion of shallow detail features and deep semantic features; The anchor frame size is recalculated based on K-means clustering to adapt to the vehicle scale distribution in haze scenes.

[0008] A vehicle target detection system in a hazy environment based on improved YOLO11, including: Data preprocessing module, used to collect haze images and construct data sets; Defogging enhancement module, which realizes the joint defogging enhancement of improved AOD-Net and Zero-DCE; Detection model module, including an improved YOLO11 network to achieve vehicle feature extraction and detection; The loss function optimization module combines SSIM and CIoU loss to optimize model training.

[0009] Preferably, the defogging enhancement module includes: AOD-Net submodule, including skip connections and SE-PA attention mechanism; The Zero-DCE submodule achieves low-light enhancement through a deep curve estimation network.

[0010] Compared with the prior art, the present invention has the following beneficial effects: Through the self-attention mechanism, the present invention can adaptively ignore irrelevant background information when the vehicle is mixed with the surrounding background, thereby improving the vehicle's recognition, reducing false detection or missed detection in the case of occlusion, and accelerating the convergence of the detection model by optimizing the training loss function.

[0011] First, skip connections are incorporated into the basic downsampling module of AOD-Net, significantly improving the network's ability to capture detailed image features. At the same time, in the intermediate feature fusion stage, shallow and deep features are fused at multiple scales to fully exploit the complementary information of features at different levels. Secondly, a new channel attention mechanism consisting of a SE attention module and a PA attention module in parallel is introduced in the decoding stage to accurately focus on key image areas. Finally, a composite evaluation index that integrates image brightness, contrast, and structural similarity is designed as a loss function. The Zero-DCE model is used as the low-light enhancement module and integrated into the AOD-Net with the improved loss function to achieve better dehazing and enhancement effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the process framework of the present invention; Figure 2 This is a vehicle target detection framework diagram of the present invention; Figure 3 This is the AOD-Net defogging model diagram of the present invention; Figure 4 This is the Zero-DCE framework diagram of the present invention; Figure 5 This is the improved defogging enhancement network diagram of the present invention. DETAILED DESCRIPTION

[0013] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0014] Example 1: A vehicle target detection method in a hazy environment based on an improved YOLO11 includes the following steps: Perform dehazing and enhancement preprocessing on haze images, and generate clear images by improving the network combining AOD-Net and Zero-DCE; For the input haze image , by improving the network combining AOD-Net and Zero-DCE to generate dehazed enhanced images The specific process is as follows: Estimating fusion parameters using improved AOD-Net , through the formula Generate a preliminary defogging image, where b is a constant deviation with a value of 1. The calculation formula is: in, is the atmospheric projection rate, A is the atmospheric illumination value; The Zero-DCE module is introduced for low-light enhancement, and the DCE-Net output curve parameters are estimated through the depth curve. , adjust the image dynamic range pixel by pixel, and the enhanced image expression is: in, is the enhancement function based on the curve parameters, n is the number of iterations; The defogging enhancement preprocessing includes: Introducing skip connection and SE-PA parallel channel attention mechanism in AOD-Net; Use the Zero-DCE module for low-light enhancement and adjust the image dynamic range through the depth curve estimation network; A composite loss function including SSIM, brightness and contrast is used to optimize the dehazing process.

[0015] The preprocessed image is input into the improved YOLO11 network, and the SE attention mechanism is used to enhance vehicle feature extraction; The feature fusion module of the improved YOLO11 network includes: Perform channel attention weighted fusion of shallow detail features and deep semantic features; The anchor frame size is recalculated based on K-means clustering to adapt to the vehicle scale distribution in haze scenes.

[0016] Adopting an adaptive multi-scale feature fusion strategy to integrate vehicle features of different scales; Perform target box regression based on optimized anchor box and CIoU loss, and output detection results.

[0017] A vehicle target detection system in a hazy environment based on an improved YOLO11 is characterized by comprising: Data preprocessing module, used to collect haze images and construct data sets; Defogging enhancement module, which realizes the joint defogging enhancement of improved AOD-Net and Zero-DCE; The defogging enhancement module includes: AOD-Net submodule, including skip connections and SE-PA attention mechanism; Zero-DCE submodule, which achieves low-light enhancement through a deep curve estimation network; Detection model module, including an improved YOLO11 network to achieve vehicle feature extraction and detection; The loss function optimization module combines SSIM and CIoU loss to optimize model training. Specific embodiment 1: (1) Data collection and preprocessing: First, a dataset specifically designed for vehicle detection in hazy conditions was constructed. By capturing images in the field and collecting publicly available videos of hazy scenes, a large amount of image data containing vehicles was obtained. These images covered varying haze concentrations, lighting conditions, and vehicle types and postures. The raw data was annotated to accurately identify the location and category of each vehicle. Considering the impact of haze on image quality, image enhancement techniques were used to preprocess the data. Methods such as histogram equalization and contrast-limited adaptive histogram equalization (CLAHE) were used to enhance image contrast and improve the distinction between vehicle targets and backgrounds. Furthermore, image dehazing algorithms, such as the dark channel prior dehazing algorithm, were used to remove haze interference from the images, making vehicle outlines clearer and providing high-quality data for subsequent model training.

[0019] (2) Research on vehicle target detection model based on improved YOLO11: This paper will design a vehicle target detection model based on the improved YOLO11, starting from data collection, preprocessing, data segmentation, feature extraction, model construction and system evaluation, to achieve improved recognition accuracy and generalization. The framework diagram of vehicle target detection designed in this paper is as follows: Figure 2 .

[0020] To address the difficulties of vehicle detection in hazy conditions, the YOLO11 model was improved. In the backbone network, the Squeeze-and-Excitation (SE) attention mechanism was introduced, enabling the model to focus more closely on the vehicle target area, suppressing background noise interference and improving feature extraction accuracy. In the feature fusion process, the Feature Pyramid Network (FPN) architecture was improved, optimizing the fusion of feature maps of different scales to better adapt to the large scale variations of vehicle targets in hazy conditions and enhance the detection of small vehicles. Furthermore, the model's anchor box settings were adjusted, and the anchor box sizes were recalculated on the haze dataset using the K-means clustering algorithm to better reflect the actual size distribution of vehicles in hazy scenes, thereby improving detection accuracy.

[0021] (3) Dehazing algorithm based on improved AOD-Net: The scenario targeted by this invention is vehicle detection in foggy and hazy weather, which places high demands on the real-time and lightweight nature of the defogging algorithm. Therefore, this invention ultimately adopts AOD-Net, which is smaller and has better real-time performance. AOD-Net is a lightweight defogging network based on deep learning. This algorithm also relies on the atmospheric defogging model, but the difference is that AOD-Net fuses the atmospheric transmittance t(x) and the atmospheric light value A into a single K(x). After integration, according to formula 1-1, it can be inferred that the converted fog-free image output is as follows: Where: K(x) can be derived as: Where: b in the formula is a constant deviation, and its value is 1.

[0022] AOD-Net integrates the atmospheric illumination value and atmospheric transmittance through a formula, making the conversion relationship between foggy images and fog-free images simpler and reducing the error. Now, we only need to know the value of K(x) to generate a clear image. Therefore, the central idea of ​​AOD-Net is to establish an adaptive depth model and use a convolutional neural network to estimate the K value of the input foggy image. After obtaining the K value, the clear image generation module can generate a new fog-free image according to the converted formula. The AOD-Net model structure is as follows Figure 3 shown.

[0023] In low-light conditions, the AOD-Net algorithm performs poorly, and even causes further reduction in image clarity and brightness. For this reason, this article hopes to introduce a low-light enhancement module to improve this problem. The Zero-DCE model is used as the low-light enhancement module. Zero-DCE is a zero-reference depth estimation curve for low-light enhancement. The algorithm uses a lightweight DCE-Net to estimate pixels and high-order curves, and determines the dynamic range adjustment of the image through the estimated values. At the same time, certain improvements have been made to the curve estimation, which can better preserve the contrast of adjacent pixels. Compared with other algorithms, Zero-DCE pays more attention to pixel range, monotonicity, and differentiability, and thereby achieves better pixel-level dynamic adjustment. This algorithm better maintains the original color of the picture while also reducing overexposure. The Zero-DCE framework is as follows Figure 4 As shown: As can be seen from the framework diagram, the Zero-DCE framework mainly consists of three parts: deep curve estimation network (DCE-Net), image enhancement module and no-reference loss function.

[0024] DCE-Net is a core component, a lightweight convolutional neural network that learns the mapping between low-light images and curve parameter maps. Its simple structure, typically consisting of multiple convolutional layers, effectively extracts image features and outputs parameters describing the enhancement curve for each pixel. For example, when processing low-light landscape images, it can accurately calculate parameters suitable for each region.

[0025] The image enhancement module enhances images based on the parameters output by DCE-Net. It uses parameterized curve functions to adjust visual attributes such as brightness and contrast pixel by pixel. After multiple iterations, it significantly improves image quality, making previously dark indoor photos clear and bright, and even rendering dark objects clearly.

[0026] The no-reference loss function is used to train DCE-Net without a reference image. It evaluates image quality based on multiple aspects, including spatial consistency, exposure control, color constancy, and illumination smoothness. By minimizing these losses, DCE-Net learns optimal enhancement strategies, ensuring that enhanced images appear natural and clear, while avoiding over- or under-enhancement.

[0027] The original AOD-Net loss function is replaced, and the Zero-DCE model is used as the low-light enhancement module, which is integrated into the AOD-Net with the improved loss function to obtain better defogging and enhancement effects. The specific network construction ideas are as follows: Figure 5 shown.

[0028] (4) Model training and optimization: A transfer learning strategy was employed, initializing the model using parameters from a YOLO11 model pre-trained on a large-scale general-purpose image dataset. This model was then fine-tuned using a constructed dataset of vehicles in hazy weather. During training, an appropriate loss function was selected, combining classification, regression, and confidence losses, and weights were adjusted to balance the various losses. Stochastic gradient descent (SGD) and its variants were employed, with dynamic learning rate adjustments to accelerate model convergence and prevent overfitting. Furthermore, by setting appropriate training hyperparameters, the model was fully trained to continuously learn the characteristic patterns of vehicle targets in hazy weather.

[0029] After training, the improved YOLO11 model was evaluated using an independent test dataset. Metrics such as mean average precision (mAP), recall, and precision were used to comprehensively measure the model's performance in detecting vehicle targets in foggy and hazy conditions.

[0030] (5) Experimental plan: The operating system used was Windows 10, equipped with an NVIDIA GeForce RTX 3080 Ti GPU with 12GB of video memory. The processor was an Intel(R) Core(TM) i7-10700 2.90GHz with 64GB of RAM. This experiment was based on CUDA 11.7 and PyTorch 1.13, using Python 3.8 as the programming language and PyCharm as the compiler.

[0031] In terms of data collection, we used the KITTI dataset for computer vision tasks to collect vehicle data. However, many of the vehicles in the dataset were obscured or difficult to identify due to weather and lighting conditions. Therefore, we needed to manually select and process the images in the dataset, annotate the vehicles using labelimg, and generate the corresponding XML file for testing using the YOLO11 model.

[0032] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A vehicle target detection method in a hazy environment based on improved YOLO11, characterized in that: The following steps are involved: Perform dehazing and enhancement preprocessing on haze images, and generate clear images by improving the network combining AOD-Net and Zero-DCE; The preprocessed image is input into the improved YOLO11 network, and the SE attention mechanism is used to enhance vehicle feature extraction; Adopting an adaptive multi-scale feature fusion strategy to integrate vehicle features of different scales; Perform target box regression based on optimized anchor box and CIoU loss, and output detection results.

2. The method according to claim 1, characterized in that For the input haze image , by improving the network combining AOD-Net and Zero-DCE to generate dehazed enhanced images The specific process is as follows: Estimating fusion parameters using improved AOD-Net , through the formula Generate a preliminary defogging image, where b is a constant deviation with a value of 1. The calculation formula is: in, is the atmospheric projection rate, A is the atmospheric illumination value; The Zero-DCE module is introduced for low-light enhancement, and the DCE-Net output curve parameters are estimated through the depth curve. , adjust the image dynamic range pixel by pixel, and the enhanced image expression is: in, is the enhancement function based on the curve parameters, and n is the number of iterations.

3. The method according to claim 1, characterized in that The defogging enhancement preprocessing includes: Introducing skip connection and SE-PA parallel channel attention mechanism in AOD-Net; Use the Zero-DCE module for low-light enhancement and adjust the image dynamic range through the depth curve estimation network; A composite loss function including SSIM, brightness and contrast is used to optimize the dehazing process.

4. The method according to claim 1, wherein The feature fusion module of the improved YOLO11 network includes: Perform channel attention weighted fusion of shallow detail features and deep semantic features; The anchor frame size is recalculated based on K-means clustering to adapt to the vehicle scale distribution in haze scenes.

5. A vehicle target detection system in a hazy environment based on improved YOLO11, characterized in that: include: Data preprocessing module, used to collect haze images and construct data sets; Defogging enhancement module, which realizes the joint defogging enhancement of improved AOD-Net and Zero-DCE; Detection model module, including an improved YOLO11 network to achieve vehicle feature extraction and detection; The loss function optimization module combines SSIM and CIoU loss to optimize model training.

6. The system according to claim 5, characterized in that The defogging enhancement module includes: AOD-Net submodule, including skip connections and SE-PA attention mechanism; The Zero-DCE submodule achieves low-light enhancement through a deep curve estimation network.

Images