FogGuard：提升YOLO在雾天的物体检测准确性

版权申诉

114 浏览量更新于2024-08-03 1 收藏 530KB PDF 举报

"FogGuard是一种针对YOLO的新型雾感知物体检测网络，旨在解决雾天对自动驾驶系统中深度神经网络（DNN）性能的影响。它采用YOLOv3作为基础，通过引入感知损失算法实现对雾气环境的补偿，以确保在恶劣天气下也能进行高精度的目标检测。在PASCAL VOC和RTTS数据集上的测试显示，FogGuard的性能优于YOLOv3，尤其是在RTTS数据集上实现了69.43%的mAP，而YOLOv3为57.78%。尽管训练过程复杂度增加，但在推理阶段FogGuard并未增加额外的时间成本。" 在当前的自动驾驶技术中，精确的物体检测是至关重要的，但雾气等恶劣天气会降低DNN的性能。传统的解决方案主要分为两类：图像增强和域适应方法。图像增强试图通过生成无雾图像来改善检测，但这通常比直接在有雾图像中检测物体更为困难。另一方面，域适应方法不依赖目标领域的标记数据，也未能完全解决问题。FogGuard则采用不同的策略，它基于微调，设计了一个专门针对雾天条件的框架。 FogGuard的核心创新在于其使用了感知损失算法，这是一种新颖的师生学习方法。在师生学习中，教师网络（在这种情况下是对清晰图像有经验的网络）指导学生网络（即FogGuard）学习在雾天环境中检测目标。感知损失关注于图像的高级特征，帮助网络理解在雾中物体的视觉表示，从而提高检测的准确性。实验结果显示，FogGuard在网络性能上有显著提升。在PASCAL VOC和RTTS这两个常用的数据集上进行了广泛的评估，特别是在RTTS数据集上，FogGuard的平均精度（mAP）达到了69.43%，相比YOLOv3的57.78%有了显著的提升。这证明了FogGuard在雾天条件下具有更强的物体检测能力。尽管在训练阶段，FogGuard引入了更高的计算复杂度，但在实际应用时，即推理阶段，它的运行效率并没有下降，这意味着FogGuard可以在保持高效的同时提供更好的雾天检测性能。这对于需要实时响应的自动驾驶系统来说，是一个重要的优势。 FogGuard通过感知损失算法为YOLO模型提供了雾天环境下的增强，提升了自动驾驶系统在恶劣天气条件下的可靠性和安全性。这种技术的出现，对于推动自动驾驶领域的发展，尤其是在应对极端天气挑战方面，具有重大的理论和实践意义。

FogGuard: guarding YOLO against fog using perceptual loss

Soheil Gharatappeh, Sepideh Neshatfar, Salimeh Yasaei Sekeh

and Vikas Dhiman

Abstract— In this paper, we present a novel fog-aware object

detection network called FogGuard, designed to address the

challenges posed by foggy weather conditions. Autonomous

driving systems heavily rely on accurate object detection algo-

rithms, but adverse weather conditions can signiﬁcantly impact

the reliability of deep neural networks (DNNs).

Existing approaches fall into two main categories, 1) image

enhancement such as IA-YOLO 2) domain adaptation based

approaches. Image enhancement based techniques attempt to

generate fog-free image. However, retrieving a fogless image

from a foggy image is a much harder problem than detecting

objects in a foggy image. Domain-adaptation based approaches,

on the other hand, do not make use of labelled datasets in the

target domain. Both categories of approaches are attempting

to solve a harder version of the problem. Our approach builds

over ﬁne-tuning on the

Our framework is speciﬁcally designed to compensate for

foggy conditions present in the scene, ensuring robust perfor-

mance even. We adopt YOLOv3 as the baseline object detection

algorithm and introduce a novel Teacher-Student Perceptual

loss, to high accuracy object detection in foggy images.

Through extensive evaluations on common datasets such as

PASCAL VOC and RTTS, we demonstrate the improvement

in performance achieved by our network. We demonstrate that

FogGuard achieves 69.43% mAP, as compared to 57.78% for

YOLOv3 on the RTTS dataset.

Furthermore, we show that while our training method

increases time complexity, it does not introduce any additional

overhead during inference compared to the regular YOLO

network.

I. INTRODUCTION

Adverse weather conditions such as rain, snow, and fog

present risks for driving. One such risk is reduced visibility,

which, in autonomous driving, impairs object detection. This

is highly dangerous; objects that are not spotted cannot

be avoided, while objects that are inaccurately localized or

classiﬁed can cause the vehicle to respond by swerving or

“phantom braking” [1]. In this work, we focus on improving

object detection in foggy weather.

We focus on improving object detection accuracy using

only cameras. Not all autonomous vehicles have multiple

sensor types, but cameras are present on virtually all of

them [2], [3]. This makes our research widely applicable,

including to vehicles that have additional sensor types;

camera-based object detection can always be combined with

other systems to improve overall accuracy via multi-sensor

fusion [4]. Other research has explored the use of fog-speciﬁc

School of Computing and Information Science, University of Maine,

Orono, ME, United States, soheil.gharatappeh@maine.edu

Department of Electrical and Computer Engineering, University of

Maine, Orono, ME, United States

This material is based upon work supported by the National Science

Foundation under Grant No 2218063

supplemental sensors, such as the novel millimeter-wave

radar [5], [6].

The image processing community has explored the prob-

lems of dehazing, defoggiﬁcation, and image-enhancement

before the success of deep learning based approaches [7]–

[10]. Bringing image processing based approaches into the

learning domain, IA-YOLO [11] combines an image pro-

cessing module with a learning pipeline to infer a de-fogged

image before feeding it into a regular object detector like

YOLO [12]. We posit that inferring a de-fogged image is

a much harder problem than detecting objects in a foggy

image. Clearly, detecting and classifying a bounding box

in a foggy image as an object class, for example car, is a

much easier problem than recreating every pixel of that car.

Additionally, dehazzing based approaches often suffer from

signiﬁcant computational overhead in order to achieve better

image quality.

To improve object detection in a foggy image, we modify

the training process of a YOLO-v3 [12] network to be robust

to foggy images. Our modiﬁed training process contributes

two novel ideas, 1) generalization of perceptual loss [13] to

Teacher-student perceptual loss (Section IV-A) and 2) data-

augmentation with depth-aware realistic fog (Section IV-B).

We use perceptual loss based on the intuition that the seman-

tic information in a foggy image is same as that of a clear

image. So we seek to minimize the perceptual loss between

a clear image and the foggiﬁed version of that iamges. Data-

augmentation is necessary because foggy object detection

datasets like RTTS [14] (∼3K images) are much smaller

than of clear image datasets like PASCAL VOC [15](∼16K

images) and MS-COCO [16] (∼116K images). Our ablation

studies demonstrate the utility of each of our contribution in

improving the accuracy of object detection in the presence

of fog.

We evaluate and compare our proposed method on

RTTS dataset against state-of-the-art approaches like IA-

YOLO [11], DE-YOLO [17] and SSD-Entropy [4]. We

ﬁnd that our approach is more accurate than IA-YOLO by

11.64% and [4] by 14.27% while being faster by a factor of

5 times.

II. RELATED WORK

Vanilla object detection algorithms [12], [18], [19] are

often insufﬁcient in adverse weather conditions such as fog,

rain, snow, low light scenarios. To address such problems,

the literature can be categorized into four main categories,

1) analytical image processing techniques, 2) learning-based

approaches, 3) domain adaptation and 4) learning-based

image-enhancement techniques.

arXiv:2403.08939v1 [cs.CV] 13 Mar 2024

下载后可阅读完整内容，剩余7页未读，立即下载

人工智能_SYBH

粉丝: 4w+
资源: 222

FogGuard：提升YOLO在雾天的物体检测准确性

Neural Architecture Search：使用Ultralytics框架进行YOLO-NAS目标检测

使用YOLO进行实时目标检测：项目实战.md

YOLO-TF2:使用TensorFlow2实现YOLO

humble-yolo:使用Keras的最小YOLO V1实现

YOLO-V5:使用对象检测模型YOLO-V5对图像进行定位和分类

yolo-unity:适用于Unity的YOLO游戏中对象检测（Windows）

go-darknet:去绑定Darknet（YOLO v4 v3）

YOLOV3损失函数解析：DarkNet框架中的yolo_layer.c详解

高铁站监控数据集：行人排队检测与Yolo标注

深度学习目标检测：R-CNN与YOLO算法解析

最新资源