Vision-based vehicle detection for a driver assistance system

On the basis of a cost-effective embedded system, this work implements a preceding vehicle detection system by using computer vision technologies. The road scenes are acquired with a monocular camera. The features of the vehicle in front are extracted and recognized by the proposed refined image processing algorithm, and a tracking process based on optical flow is also applied for reducing the complexity of computing. The system also provides the longitudinal distance information for the further function of adaptive cruise control. Moreover, voice alerts and image recording will be activated if the distance is less than the safe range. A statistical base of 100 video road images are tested in our experiments; the natures of the vehicles include sedan, minivan, truck, and bus. The experimental results show that the proportion of correct identifications of proceeding vehicles is above 95.8%, testing on highways in the daytime. Experimental results also indicate that the system correctly identifies vehicles in real time.     

Vision-based vehicle guidance

This implementation of lane and obstacle detection for an autonomous, self-guided vehicle succeeds by tailoring vision and computational techniques to an affordable SIMD architecture. The authors use a geometrical transform called inverse perspective mapping (IPM). Using a priori knowledge of both the scene and the acquisition device, the IPM technique allows one to remove the perspective effect and produce a new image in which the information content is homogeneously distributed among all pixels. In the remapped image, the amount of information carried by each pixel no longer depends on the pixel's position, making the SIMD approach practical     

引入辅助损失的多场景车道线检测

目的 为解决实时车辆驾驶中因物体遮挡、光照变化和阴影干扰等多场景环境影响造成的车道线检测实时性和准确性不佳的问题,提出一种引入辅助损失的车道线检测模型。方法 该模型改进了有效的残差分解网络（effcient residual factorized network,ERFNet）,在ERFNet的编码器之后加入车道预测分支和辅助训练分支,使得解码阶段与车道预测分支、辅助训练分支并列,并且在辅助训练分支的卷积层之后,利用双线性插值来匹配输入图像的分辨率,从而对4条车道线和图像背景进行分类。通过计算辅助损失,将辅助损失以一定的权重协同语义分割损失、车道预测损失进行反向传播,较好地解决了梯度消失问题。语义分割得到每条车道线的概率分布图,分别在每条车道线的概率分布图上按行找出概率大于特定阈值的最大点的坐标,并按一定规则选取相应的坐标点,形成拟合的车道线。结果 经过在CULane公共数据集上实验测试,模型在正常场景的F1指标为91.85%,与空间卷积神经网络（spatial convolutional neural network,SCNN）模型相比,提高了1.25%,比其他场景分别提高了1%~7%;9种场景的F1平均值为73.76%,比目前最好的残差网络——101-自注意力蒸馏（ResNet-101-self attention distillation,R-101-SAD）模型（71.80%）高出1.96%。在单个GPU上测试,每幅图像的平均运行时间缩短至原来的1/13,模型的参数量减少至原来的1/10。与平均运行时间最短的车道线检测模型ENet——自注意力蒸馏（ENet-self attention distillation,ENet-SAD）相比,单幅图像的平均运行时间减短了2.3 ms。结论 在物体遮挡、光照变化、阴影干扰等多种复杂场景下,对于实时驾驶车辆而言,本文模型具有准确性高和实时性好等特点。     

车道模型的高速公路车辆异常行为检测方法

针对高速公路上车辆的逆行、停车、轨迹异常等事件的检测问题,提出了一种基于车道模型知识的自底上向的车辆异常检测方法。首先由车道线的连续性、共线性的感知搜索出车道线和消失点,自动建立车道模型;然后在多车辆检测与跟踪时,通过目标运动位置预测和KLT特征点跟踪的方法建立表示目标区域交叠关系图,依据后验关系通过对图中节点对应目标区域的合并与拆分实现目标与实际车辆的一一对应,建立可靠的跟踪轨迹;最后基于消失点的坐标变换,计算车辆实际位置与速度,采用轨迹滑动窗口方法判断目标运动趋势并计算窗口内轨迹点平均速度,判断车辆的异常行为。实验结果表明,所提方法在不同天气、不同车流量环境中均有80%以上的事件检测率,同时算法简单,具有很好的实时性,能够适应实际高速公路智能检测设备的需求。     

基于实例分割的车道线检测及自适应拟合算法

车道线检测是智能驾驶系统的重要组成部分。传统车道线检测方法高度依赖手动选取特征,工作量大,在受到物体遮挡、光照变化和磨损等复杂场景的干扰时精度不高,因此设计一个鲁棒的检测算法面临着很大挑战。为了克服这些缺点,提出了一种基于深度学习实例分割方法的车道线检测模型。该模型基于改进的Mask R-CNN模型,首先利用实例分割模型对道路图像进行分割,提高车道特征信息的检测能力;然后使用聚类模型提取离散的车道线特征信息点;最后提出一种自适应拟合的方法,结合直线和多项式两种拟合方法对不同视野内的特征点进行拟合,生成最优车道线参数方程。实验结果表明,该方法提高了检测速度,在不同场景下都具有较好的检测精度,能够实现对各种复杂实际条件下的车道线信息的鲁棒提取。     

Vision-based navigation of an unmanned surface vehicle with object detection and tracking abilities

The paper discusses autocalibration, object detection, and object tracking for unmanned surface vehicles. Input data are recorded with a wide-baseline stereo vision system providing accuracy for distance estimations. The paper reports about followed ways and novel contributions for ensuring a working system solution. Automatic self-calibration is used for the wide-baseline stereo vision system. Robust sea surface estimation and the detection of the horizon support the understanding of the given scene environment. Long-range (i.e. up to 500 m) object detection and tracking are supported by the used wide-baseline stereo system. The paper informs about the complete system design, informs about applied or designed methods, and also about experiments which verify that the system achieved an operational state.     

Adaptive lane line detection and warning algorithm based on dynamic constraint

Multimedia Tools and Applications - The lane line detection algorithm is highly sensitive to the environment. And the selection of parameters is greatly affected by human subjective factors....     

基于密集分割网络的车道线检测方法

传统车道线检测算法大多数依赖手工制作特征和启发式算法的组合,容易受车辆遮挡和地面污损等因素的影响。针对影响车道线检测的复杂问题,将车道线检测视为连续细长区域实例分割问题,提出了一种基于密集分割网络的车道线检测方法。为此,使用稠密块构建了一个密集分割网络DSNet,该网络能够利用特征重复使用的特性提高提取车道线实例特征和恢复特征图分辨率的性能。同时,还引入了邻近AND运算和Meanshift聚类算法对DSNet网络的输出进行处理,减小了非车道线像素的影响,使得检测结果的边界线更为清晰。实验表明,本文方法能很好地解决车辆遮挡和地面污损问题,并且还能确定车道线的数量,具有较好的鲁棒性和实时性。     

一种基于视觉的道路检测算法

计算机视觉导航使移动机器人可以工作于复杂环境中,而道路检测是其中的关键环节。针对道路的多样性和环境因素的影响,提出了一种稳健的道路检测算法。依据道路特征把图像分为3组区域：路、非路和不确定区域,然后对难以判断的不确定区域使用假设检验的策略,依据道路的形状、路宽和面积信息综合判断,把不确定区域合并到路或者非路区域,从而快速准确地检测出道路。该算法已经在移动机器人-ATRV上测试和使用。     

Vision-based autonomous land vehicle guidance in outdoor road environments using combined line and road following techniques

An intelligent approach to autonomous land vehicle (ALV) guidance in outdoor road environments using combined line and road following and color information clustering techniques is proposed. Path lines and road boundaries are selected as reference models, called the line-model and the road-model, respectively. They are used to perform line-model matching (LMM) and road-model matching (RMM) to locate the ALV for line and road following, respectively. If there are path lines in the road, the LMM process is used to locate the ALV because it is faster than the RMM process. On the other hand, if no line can be found in the road, the RMM process is used. To detect path lines in a road image, the Hough transform is employed, which does not take much computing time because bright pixels in the road are very few. Various color information on roads is used for extracting path lines and road surfaces. And the ISODATA clustering algorithm based on an initial-center-choosing technique is employed to solve the problem caused by great changes of intensity in navigations. The double model matching procedure combined with the color information clustering process enables the ALV to navigate smoothly in roads even if there are shadows, cars, people, or degraded regions on roadsides. Some intelligent methods to speed up the model matching processes and the Hough transform based on the feedback of the previous image information are also presented. Successful navigations show that the proposed approach is effective for ALV guidance in common roads. ©1997 John Wiley & Sons, Inc.     

Vision-based global localization and mapping for mobile robots

We have previously developed a mobile robot system which uses scale-invariant visual landmarks to localize and simultaneously build three-dimensional (3-D) maps of unmodified environments. In this paper, we examine global localization, where the robot localizes itself globally, without any prior location estimate. This is achieved by matching distinctive visual landmarks in the current frame to a database map. A Hough transform approach and a RANSAC approach for global localization are compared, showing that RANSAC is much more efficient for matching specific features, but much worse for matching nonspecific features. Moreover, robust global localization can be achieved by matching a small submap of the local region built from multiple frames. This submap alignment algorithm for global localization can be applied to map building, which can be regarded as alignment of multiple 3-D submaps. A global minimization procedure is carried out using the loop closure constraint to avoid the effects of slippage and drift accumulation. Landmark uncertainty is taken into account in the submap alignment and the global minimization process. Experiments show that global localization can be achieved accurately using the scale-invariant landmarks. Our approach of pairwise submap alignment with backward correction in a consistent manner produces a better global 3-D map.     

Vision-based entrance detection in outdoor scenes

Doors are a significant object for the visually impaired and robots to enter and exit buildings. Although the accuracy of door detection is reported high in indoor scenes, it has become a difficult problem in outdoor scenes in computer vision. The reason may lie in the fact that such properties of a simple ordinary door such as handles, corners, and the gap between the door and the ground may not be visible due to the great variety of doors in outdoor environments. In this paper, we present a vision-based method for detecting building entrances in outdoor images. After extracting the lines and deleting the extra ones, regions between the vertical lines are specified and the features including height, width, location, color, texture and the number of lines inside the regions are obtained. Finally, some additional knowledge such as door existence at the bottom of the image, a reasonable height and width of a door, the difference between color and texture of the doors and those of the neighboring regions, and numerous lines on doors is used to decide on door detection. The method was tested on the eTRIMS dataset, door images from the ImageNet dataset, and our own dataset including doors of houses, apartments, and stores leading to acceptable results. The obtained results show that our approach outperforms comparable state-of-the-art approaches.     

一种基于线扫描相机的新型交通检测器

针对基于视频的交通检测器的检测精度和实时性难以提高的问题,提出了一种基于线扫描相机的新型交通检测器。该检测器通过两个相距2米的线扫描相机捕获移动车辆的俯视图像,采用小波变换算法进行图像背景更新和目标提取,运用目标投影曲线相关匹配的方法进行车辆速度估计,最终检测出经过检测断面所有车辆的个数和速度。实验结果表明,基于线扫描相机的交通检测器在车辆计数和车速估计方面的性能上均要优于传统的基于视频的交通检测器。     

Robust and ubiquitous smartphone-based lane detection

Lane-level positioning is required for several location-based services such as advanced driver assistance systems, driverless cars, predicting driver’s intent, among many other emerging applications. Yet, current outdoor localization techniques fail to provide the required accuracy for estimating the car’s lane.In this paper, we present LaneQuest: an accurate and energy-efficient smartphone-based lane detection system. LaneQuest leverages hints from the ubiquitous and low-power inertial sensors available in commodity off-the-shelf smartphones about the car’s motion and its surrounding environment to provide an accurate estimate of the car’s current lane position. For example, a car making a u-turn, most probably, will be in the left-most lane; a car passing by a pothole will be in the pothole’s lane; and the car angular velocity when driving through a curve reflects its lane. Our investigation shows that there are amble opportunities in the environment, i.e. lane “anchors”, that provide cues about the car lane. To handle the ambiguous location, sensors noise, and fuzzy lane anchors; LaneQuest employs a novel probabilistic lane estimation algorithm. Furthermore, it uses an unsupervised crowd-sourcing approach to learn the position and lane span distribution of the different lane-level anchors.Our evaluation results from implementation on different Android devices and driving traces in different cities covering 260 km shows that LaneQuest can detect the different lane-level landmarks with an average precision and recall of more than 91%. This leads to an accurate detection of the exact car lane position 84% of the time, increasing to 92% of the time to within one lane. This comes with a low-energy footprint, allowing LaneQuest to be implemented on the energy-constrained mobile devices.     

Vision-based obstacle detection for rotorcraft flight

The ability of rotorcraft to fly at low altitude is hindered by the high pilot workload required to avoid obstacles. The development of automation tools that can detect obstacles in the rotorcraft flight path, warn the crew, and interact with the guidance system to avoid detected obstacles would significantly reduce pilot workload and increase safety. This article describes an obstacle detection approach based on feature tracking and recursive range estimation that takes into account the characteristics of rotorcraft flight. The merits and weaknesses of the approach are discussed using image sequences from the laboratory and from flight. © 1992 John Wiley & Sons, Inc.     

高速公路弯道边界的识别与重建

分析了智能车辆安全辅助驾驶系统中弯道分道线的检测提取方法,提出一种基于道路区域分割的弯道检测新算法,包含道路区域分割和弯道边界检测。在分割出道路区域和天空区域并划定弯道检测的感兴趣区域后,提取分道线候选点,并对候选点进行校正,最终拟合并重建出弯道分道线,且准确判断了车道线弯曲方向。实验证明,该算法的实时性和准确性均高于在整幅图像中提取车道线的传统方法。     

基于直线段匹配的移动机器人的障碍物检测

阐述了基于机器视觉的前方障碍物的检测问题。提出一种基于直线段提取和匹配并最终确定障碍物的轮廓的方法。试验表明该方法运算速度快,结果可靠。     

智能汽车换道控制算法的仿真研究

针对传统智能车辆跟随轨迹控制方法所存在的延迟反应问题,基于预瞄一跟随理论建立了智能车辆换道过中的轨迹跟随运动模型,提出了智能车辆换道过程中的控制算法.在PreScan和matlab/simulink的联合仿真环境下,实现了智能车辆换道过程中轨迹跟随控制,并进行了36km/h、72km/h和108km/h速度下的仿真验证.仿真结果表明,仿真轨迹与实测换道轨迹走势接近且重合度较高.     

基于扫描线和特征筛选的车牌定位快速算法

以对车牌识别准确率影响最大的车牌定位技术为重点,研究并提出一种基于扫描线和特征筛选的车牌定位算法,该算法先记录并分析二值图像中相邻水平扫描线上的跳变点信息,确定出候选车牌区域,再根据车牌特征筛选,最终确定车牌区域.对113幅不同车型的图像进行测试,结果表明,去噪处理对定位准确率有很大影响,当车牌倾斜角度小于5°、且经过去噪处理时,定位准确率超过90%,定位时间小于0.9s.