Current Issue Cover
结合KSW和FCNN的道路场景分割研究

王云艳,罗冷坤,周志刚(湖北工业大学)

摘 要
随着自动驾驶技术不断引入生活,机器视觉中道路场景分割算法的研究已至关重要。传统方法中大多数研究者使用机器学习的方式来对阈值分割,近些年深度学习的引入,使得积神经网络被广泛应用于该领域。针对传统的阈值分割方法难以有效提取多场景下道路图像阈值的问题和直接用深度神经网络来训练数据导致过分割严重的问题,本文提出了结合KSW和全卷积神经网络(FCNN)的道路场景分割方法,该方法结合了KSW熵法及遗传算法,利用深度学习在不同场景下的特征提取,并将其运用到无人驾驶技术的道路分割中。首先将道路场景测试集利用KSW熵法及遗传算法得到训练集,然后导入到全卷积神经网络来训练得到有效训练模型,最后通过训练模型可实现对任意一张道路场景图分割。本文实验结果表明,在KITTI数据集中进行测试,天空和树木的分割精度分别达到91.3%和94.3%,道路、车辆、行人这几类分割进度提高了2%左右。从分割结果图中能明显看出,道路上的积水、泥潭等过分割现象有明显的改观。相比传统机器学习道路场景分割方法,该方法在一定程度上提高了分割精度,相比深度学习直接应用于道路场景分割的方法,本方法在一定程度上避免了过分割现象,提高了模型的鲁棒性。综上所述,本文提出的结合KSW和FCNN的道路场景分割算法有广泛的研究前景,有望应用于医学图像和遥感图像的处理中。
关键词
Road Scene Segmentation Study Based on KSW and FCNN

Wang Yun Yan,Luo LengKun,Zhou ZhiGang(Hubei University of Technology,Wuhan,China)

Abstract
The advent of driverless cars has become a hot topic in today''s society. The definition of driverless is to achieve a high degree of autonomous driving behavior through environmental awareness: start, brake, lane line tracking, lane change, collision avoidance, parking, etc. Image segmentation of road scenes plays an important role in this technology. It is of great significance to study how to achieve complex scenes and high-efficiency scene segmentation images in the environment of severe noise interference.Traditional road segmentation often uses two types of methods: a method based on binocular stereo vision and a method based on motion indicators. For example, Chen Shuangyu proposed pedestrian detection based on binocular stereo vision and SVM algorithm, and used threshold segmentation to determine the coordinate position of the moving target. For the diversity of motion indicators, Sturgess used the projection surface direction and object. Multiple motion indicators such as altitude and feature tracking density segment the road. However, the above methods have high requirements on computing resources, and it is difficult to meet the practical requirements of driverless cars at this stage. Since 2012, deep learning has been gradually introduced into road scene segmentation. Zou Bin proposed a smart car steering study based on end-to-end depth learning, and obtained good road feature coding through pre-training self-encoding; In recent years, due to the implementation of GPU parallel computing in computers, the speed of the training process for large-scale data has been greatly improved. The Convolutional Neural Network (CNN) has become a research hotspot and has been widely used, Alvarez Based on the deep learning algorithm of Convolutional Neural Network (CNN) to learn high-order features in the scene to achieve road scene segmentation, although the computational strength is reduced to some extent, there are still some problems of over-segmentation of complex scenes. Wang Hai proposed an automatic feature extraction feature using deep convolutional neural network (DCNN) depth structure for complex scene problems, supplemented by feature self-encoder to measure feature similarity in source-target scenarios. However, these algorithms do not achieve the desired results for the road marking, vehicle, and pedestrian segmentation accuracy. For the rainy days, snowy days, and high temperature weather, the road surface often appears to be divided.Aiming at the problem that the traditional threshold segmentation method is difficult to extract the road image threshold in multiple scenes effectively and the training of data directly using deep neural network leads to serious over-segmentation, this paper proposes a road scene combining KSW and full convolutional neural network (FCNN). The segmentation method, which combines the KSW entropy method and the genetic algorithm, uses depth learning to extract features in different scenarios and applies it to the road segmentation of unmanned technology. Firstly, the road scene test set is obtained by KSW entropy method and genetic algorithm, and then imported into the full convolutional neural network to train the effective training model. Finally, the training model can be used to segment any road scene map. The experimental results in this paper show that the segmentation accuracy of the sky and trees reached 91.3% and 94.3% respectively in the KITTI dataset, and the segmentation progress of roads, vehicles and pedestrians increased by about 2%. It can be clearly seen from the segmentation result map that the over-segmentation of water accumulation and mire on the road has been significantly improved. Compared with the traditional machine learning road scene segmentation method, this method improves the segmentation accuracy to a certain extent. Compared with the depth learning method, it is directly applied to the road scene segmentation method. This method avoids the over-fitting phenomenon to some extent and improves the model. Robustness. In summary, the road scene segmentation algorithm proposed in this paper combined with KSW and FCNN has broad research prospects and is expected to be applied to the processing of medical images and remote sensing images. In view of the high efficiency, accuracy and robustness of the road segmentation algorithm, this paper introduces Fully Convolutional Neural Networks (FCNN) to solve road image segmentation in complex scenes, based on deep learning. Based on the theory, the KSW entropy method and genetic algorithm are combined to improve the anti-interference of harsh environment, complex scenes and noise.
Keywords
QQ在线


订阅号|日报