0　Introduction

With the rapid increase of vehicle amount, traffic accidents have become serious problems in China. It was reported that in China 274 persons died from abnormal driving in 2015, which accounted for 8.9% of the death in traffic accidents. According to the information provided by the traffic management bureau, about 40% fateful traffic accidents result from abnormal driving such as fatigue driving and illegal lane change. Thus, the research on the abnormal driving behavior identification (ADBI) of drivers is of very important practical significance.

一要抓好畜牧生产源头治理，大力推进健康养殖,加大无公害畜产品产地和产品认证工作，鼓励规模养殖场、养殖小区、畜产品加工企业申报无公害畜产品产地和认证,大力推广无公害畜禽产品，组织开展优质畜产品进学校、进部队等活动。二要按照《动物检疫管理办法》的要求，全面加强产地检疫、屠宰检疫和市场监管工作，实现产地检疫面、屠宰检疫率达100%，畜禽交易市场、畜产品交易市场监督面达100%，严禁未经检疫或检疫不合格的畜禽及其产品进入市场。三要严格按照农业部规范实施检疫、使用检疫证章标识，严禁违规操作、违规出证。严厉打击经营病死动物及其产品的行为，保障人民群众吃上“放心肉”。

Currently, there are mainly three kinds of ways to identify abnormal driving behaviors: (1) identification based on physiological characteristics. (2) identification based on facial characteristics. (3) identification based on vehicle behavior characteristics. The identification based on physiological characteristic detects drivers’ electroencephalogram (EEG) and electrocardiograph (ECG) directly[1，2] has high reliability. However, the detection equipment used in this method is directly connected to drivers’ skin, which will more or less disturb drivers during the long driving period. The identification based on facial characteristic extracts the drivers’ facial characteristics in real time by a camera, such as blink frequency[3], to analyze driver’s abnormal driving behavior. Nevertheless, this kind of method has high error rate when a driver wears a pair of glasses[4]. The identification based on vehicle behavior characteristics detects driver’s driving behavior according to the output performance of the operated vehicle, such as the variation of steering wheel angle[5，6]. By using this method, some interference such as illumination change can be avoided, but the method is easily affected by driver’s driving habit. In conclusion, although these methods above can achieve the goal of identifying driver’s abnormal driving behavior, they still have some limitations in terms of accuracy and reliability under complex road environment.

In order to deal with the limitations of the existing methods, a new ADBI method based on lane and vanishing point detection is proposed and the lane image is preprocessed firstly by using a gradient enhancement method based on linear discriminate analysis (LDA) and the edge detection algorithm based on Sobel operator and adaptive threshold segmentation to enhance the lane edge. After selecting a region of interest, the lane is extracted by a Hough transform based on local research scope by setting constraint to polar radius and polar angle, which can improve the efficiency of the lane extraction. Then, an algorithm based on a scanning algorithm according to the brightness differences between the lanes and the roads is used to further verify the detected lanes. Finally, a new ADBI method based on the offset angle and lateral distance is proposed to judge whether the abnormal driving behavior occurs. The whole process of the ADBI is shown in Fig.1.

Fig.1　The flow chart of ADBI

The contributions of the paper are as follows. (i) A new ADBI method is proposed, where both direction and position offsets are together considered as identification condition. Compared with state-of-the-art methods, which typically consider only one aspect of direction and position offsets during ADBI, the proposed method can improve the accuracy and reliability of the ADBI due to considering two aspects of abnormal driving. (ii) A validation scheme based on priori information is proposed to further verify the detected lane, which can effectively eliminate some typical interference factors, e.g., road edge and highway barrier edge, and improve reliability of lane detection.

1　Image preprocessing

In order to filter noises and enhance the lane edge, image preprocessing should be done firstly, which will simplify the following processing steps and improve real-time performance.

1.1　Gradient enhancement based on LDA

Color is an unstable characteristic which can be easily affected by the change of illumination. If the illumination is poor, it will become difficult to distinguish the lanes from the roads. Thus, in this work, a gradient enhancement method based on LDA is proposed in order to generate optimal RGB weights and enhance the edge of lanes:

贵州省人民医院护士学校，在校院合作上一直具有得天独厚的优势。但随着办学规模的扩大，学校也面临着原有合作医院容纳能力不足，新的合作对象拓展不力，合作层次不深、领域不广，合作关系不稳定等一系列问题。如何突破传统校企合作模式的局限，实现职业教育教学“五个对接”，全面提升学校办学水平呢？对此我校拓宽校企合作思路，根据不同合作对象探索不同的合作方式，取得了比较明显的成效。

Gx=Sx×I′, Gy=Sy×I′

gl=|W·Xl-W·Xr|

(1)

gl(Wmax)=max(gl(w))

(5) By minimizing the within-class-scatter and maximizing the between-class-scatter, the lane class and the road class can be distinguished from each other. Thus,

where W is a conversion vector which converts the color image into the gradient-enhanced image, and Xl and Xr are the RGB pixel values of the lane region and the road respectively.

=max|W·Xl-W·Xr|

在行李寄存处常碰到客人取行李扯皮的事，什么东西少了、电脑坏了啊的，隔三差五就有，寄存处的人个个都练得一张铁嘴还有一张铁脸皮，他们从未败过，因此客人们只能是自认倒霉。

(2)

(3) In order to figure out Xl and Xr, training data which consist of road and lane classes are needed. Training data for class c for the frame at time t are defined as follows:

(3)

where denotes training data of class c for the tth frame at time, c represents either the lane or the road, denotes training data for class c extracted from the tth frame, and k is the number of previous frames to be used as training data for the current frame.

(4) To find Wmax in Eq.(2), the vector that can maximize discriminance between the two classes (lane class and road class) should be figured out, and this problem can be solved by LDA. LDA is used to find a linear combination of between-class scatter SB and within-class scatter SW. The definitions are as follows:

(4)

where x is data, ni is the amount of data in class i, mi denotes the mean of class i, m is the mean of all data, c is the number of classes, and Ci denotes a set of all data in class i.

(2) The key to gradient-enhancing conversion is to find vector Wmax which satisfies the following equation:

Wmax=arg max|WTSBW/WTSWW|

(5)

Tl=|Wmax·ml-Wmax·mr|

I′=Wmax·I

(6)

where I is RGB value of the original image, and I′ is the image after gradient enhancement.

1.2　Edge detection based on Sobel adaptive algorithm

After the gradient enhancement, the lanes become more noticeable and easy to be distinguished. Based on the gradient-enhanced images, the edge detection via Sobel adaptive algorithm is used to extract the edge points of the lanes for the following lane detection.

在我国经济告诉发展的现在，各种大型企业如雨后春笋般冒了出来，而对于一个大型企业而言，其管理方面自然是极为重要的。随着企业的需求不断提高，传统的管理方案已经逐渐不能满足企业的需求，必须要有更加有效的管理技术才能够支撑企业的发展。管理会计的产生，则能够很好的改善当前企业管理中存在的弊端。但是目前来说，管理会计在我国仍处于发展阶段，在企业的运用中仍然存在问题，因此如何找到管理会计在企业管理中存在的问题并解决它，就是当前企业管理发展所需要关注的重点问题。

By edge detection, edge information can be retained and the amount of data processing is reduced. Thus, self-adaptive threshold Sobel algorithm is used to extract the edge information of lanes.

The Sobel operator in x and y directions can be represented as[7]:

(7)

Assume that Gx and Gy are the horizontal and vertical gradient values of the image, then:

(1) Definition of lane gradient is as follows[7]:

(8)

And the gradient value of the pixel whose coordinates are (x, y) can be figured out:

(9)

Unlike classical Sobel algorithm which requires constant thresholds, a self-adaptive edge detection algorithm is used in the paper. The Canny edge detector uses two threshold values to determine edges. Pixels with gradient values above the larger threshold Tl are selected as edges. On the other hand, pixels with values below the smaller threshold Ts are determined as non-edges. Pixels between the two thresholds are set as edge candidates and are classified as edges if there is a path from the pixel to an edge. After gradient enhancement, the RGB vectors of the lanes and roads are normally distributed[8]. Therefore, these adaptive thresholds can be self-adaptively calculated from statistical characteristics of the road and lane class. Large threshold value Tl can be determined by

(6) Finally, the image after gray-level conversion can be calculated:

(10)

where ml and mr are the mean RGB values of the lane region and the road region respectively.

Small threshold value Ts is determined by

Ts=max(|Wmax·ml-d|, |Wmax·mr-d|)

(11)

where d is a value for which lane and road probabilities are equal[9].

2　Detection of lane and vanishing point

2.1　Selection of region of interest

Generally, useful lane information is mainly at the bottom of the road image. Thus, the bottom part of the captured image is determined as the region of interest to reduce the calculation amount of lane detection. In order to detect the two lanes in both sides of the vehicle, the region of interest is further divided into two parts, denoted as RL and RR respectively. The following steps are all implemented in the region of interest. The partition of lane image is shown in Fig.2.

Fig.2　Partition of lane image

2.2　Lane detection

Traditional Hough transform lane detection algorithms are based on a huge search scope. As a result, the calculation is time-consuming and these algorithms are not suitable for real-time online detection. Therefore, an improved Hough transform is used to improve the efficiency of the edge detection. By setting constraint to polar angle and polar radius, the search scope of the Hough transform is greatly shrunken.

Assuming that in the tth frame, the polar radiuses of the lanes on both sides of the vehicle are and and polar angles of that are and respectively. The experiments on a video involving a large amount of lane images[10,11] show that the lane in the tth frame is just adjacent to the lane in the (t+1)th frame in spatial location.

Thus, if the search scope for Hough transform is limited in the area adjacent to the lane in the tth frame before using Hough transform[11] to extract polar radius and as well as polar angle and in (t+1)th frame, the computational burden of the Hough transform will be greatly reduced, and the efficiency of the lane detection based on the Hough transform will be improved as well.

Therefore, in this paper, the lane is detected by the improved Hough transform algorithm based on local search scope. The search scope of polar radiuses and polar angles in the (t+1)th frame based on the detected polar radiuses and polar angles in the tth frame is determined as: where α and ε are two threshold values. In the paper, we let α=15, ε=10 to obtain an high-efficiency detection.

老头子还依然玩着，依然常常故意把假脚举起，作为其中一个全身均被举起的姿势，又把肩背极力倾斜向左向右，便仿佛傀儡相扑极烈。到后便依然在一种规矩中倒下，毫不苟且的倒下。自然的，王九又把赵四战胜了。

Fig.3　Constraint of lane domain

Based on the research scope determined above, we can obtain the two parameters of the detected lane using the Hough transform[7], that is, ρ and θ, where ρ and θ express polar radius and polar angle in polar coordinate system respectively. From Fig.3, the inclination angle θ′ of a line in Cartesian coordinate system can be described as

θ′=90°-θ

(12)

2.3　Validation of the detected lanes

So far, the lanes have been preliminarily detected. However, in some cases, the detected lanes could not be real ones due to the distraction from other edge lines in the image, such as the edge of the road or highway barrier. Therefore, we need to further verify the detected lanes.

这里还是“甜渣党”们的天堂，一定得去一趟那些特色的wine shop，里面可谓是“卧虎藏龙”。同行的侍酒师孙昕就偶遇一间natural wine shop，里面竟然有着3000年前米西比亚时期的墓穴，不过更让他兴奋的是那丰富、独特的自然酒和那颇有Geek范的老板。两人兴致勃勃地聊了好一会，小伙伴果断花掉身上最后一个钢，提着两箱酒开开心心地走回酒店。

An existing study[12] finds that the lane region is generally much brighter than the surrounding regions on both sides of it, so its gray-level value is obviously higher than that of the surrounding areas. Based on the priori information in structure and brightness, we formulate a set of rules to verify the detected lane. The formulated rules are as follows:

The detailed calculation process of the two parameters, direction offset and position offset, is as follows:

(13)

If there is an unreal lane across region B, such as A2B2C2 in Fig.4, then three gray averages of three windows will not satisfy the inequation (13).

Fig.4　Schematic diagram of validation of the detected lanes

(2) When the lane is blocked by other vehicles or becomes blurred,the lane might be not continuous. Therefore, it is impossible for every pixel in the real lane to satisfy in inequation (13). In order to enhance the robustness of lane validation, a constraint is set as

(14)

where NS denotes the number of pixels that satisfy the inequation(13), NT denotes the total number of pixels on the detected lane, τT is a threshold value. Based on plenty of field experiments, when τT=0.19, best detection results can be obtained.

If the grey values of pixels in the detected lane meet the constraints of inequation (13) and (14), then the detected lane is verified to be real lane. Otherwise, it is verified to be unreal.

2.4　Calculationofvanishingpoint

Assume that the parametric equation of line li is

图10所示的是图9中第3个脉搏波的处理示意图。由于噪声的影响，该脉搏波的终点幅值高于起点幅值。基线校准后，使得起点与终点的幅度相同。进而进行归一化处理，在没有影响波形形状的情况下摆脱了幅度的随机波动问题，得到一个波形完整、特征点明显的脉搏波。归一化使得幅度和面积参数更加稳定，且时间参数不受任何影响。

(15)

where is the inclination angle and ρi is the distance from the origin to the line.

Vanishing point coordinate is represented by (xM, yM). Ideally, all the lines in the image intersect at the vanishing point, namely, the sum of the distance from the vanishing point to the line is equal to zero. Thus, the calculation of the vanishing point can be transferred into an extremal problem.

An objective function is constructed in the parameter space:

(16)

The vanishing point is the point that makes the objective function obtain minimum:

董庄排闸、引闸各布置1孔，孔深15m。闸基高程5.4～8.4m主要为第②层壤土，构成地基主要持力层，具中等压缩性，弱透水性，强度较高，渗透稳定性较好。高程3.3～5.4m为第②2层砂壤土，具中等压缩性，中等透水性，具液化潜势。高程1.7～3.3m为第②3层粉砂，饱和，中密，中等透水性，承载力较高，具液化潜势。高程1.7m以下为第③层壤土，可塑～软塑，具中高等压缩性，微透水性，工程地质性质相对较差。

(xM, yM)=argminU(x, y)

(ⅰ) f-1(U)∈τ。事实上由U∈σCSI，U∈σ,又f:(X,τ)→(Y,σ)连续,于是f-1(U)∈τ。

(17)

3　Abnormal driving behavior identification

There are two causes of abnormal driving: the direction offset and position offset between the lane and the vehicle. However, the existing ADBIs only focus on one single factor and have poor robustness. Furthermore, they only give out a warning signal and cannot provide specific reason of abnormal driving when abnormal driving occurs.

随着人工智能技术的广泛应用，在给人们带来便利的同时，也必然增大犯罪发生的概率。不能只顾品味人工智能技术发展所带来的饕餮盛宴，而无视其可能带来的风险。为了避免未来人工智能时代的犯罪蔓延，有必要对犯罪预防问题进行深入探讨。

Fig.5 shows the condition when the vehicle is traveling straightly at the center of the road. MA is the detected left lane and MB is the detected right lane. M is the calculated vanishing point.

There are two cases of abnormal driving. In the first case, abnormal driving is caused by direction offset.

Fig.5　The schematic diagram of normal driving

Namely, there is an obvious angular offset between the lane and the vehicle. In the second case, the vehicle is going on a straight way, but it is too close to one side of the lane, which is also a potential threat to driving safety.

In order to quantify abnormal driving, two parameters are introduced. β denotes the direction offset between the vehicle direction and line a, which is the angular bisector of the two detected lanes, and l denotes the position offset from the middle of the image to the calculated vanishing point. From Fig.6 and Fig.7, it can be found that identifying abnormal driving behavior only by one single factor is unreliable. The reasons are as follows:

(1) In Fig.6, the thinner lines represent the detected lanes when the vehicle is traveling normally, and the thicker lines represent the detected lanes when abnormal driving caused by direction offset occurs. It is obvious that when abnormal driving caused by direction offset occurs, l can be remarkable while β is still in the normal range shown in Fig.6.

第二，新形势带来的冲击是不容忽视的，我国政府要建立在“一带一路”、“中非合作”、欧洲一体化等国际经济体系变更基础上，寻求新的国际财经合作秩序。可以欣喜的看到我国政府在此方面已经做了大量的努力与工作，现阶段所形成的国际财经合作体系已经卷入了大约80个国家与地区，形成了稳固的发展态势与规模。下一步，要尽一切可能扩大我国在国际财经合作上的成果，以实体经济作为纽带与杠杆，利用广泛应用非洲、欧洲以及东南亚的资源与市场，形成新的合作秩序，在带动第三世界国家普遍发展的同时完善自己的对外财经输出力度，形成稳步的有效市场结构。

Fig.6　Abnormal driving situation due to direction biasing

(2) When abnormal driving is caused by position offset, it is β that is in the abnormal scope, but l is within normal limits shown in Fig.7. Similarly, in Fig.7, the thinner lines represent the detected lanes when the vehicle is traveling normally, while the thicker lines represent the detected lanes when abnormal driving caused by position offset occurs.

Fig.7　Abnormal driving situation due to the close distance between the lane and the vehicle

Seen from the above description, when abnormal driving occurs one factor indicates that the vehicle has a tendency to deviate, while another factor is usually still in normal range. Thus, ADBIs which just focus on one single factor are unreliable. Also, there exists obvious difference between normal driving and abnormal driving in the direction offset and position offset. Therefore, in this study, a new ADBI method based on direction and position offsets to improve the reliability of the ADBI is proposed.

由图2可知，电磁场在离开金属表面后急剧衰减，这符合SPP的特点.其它阶模式对r的依赖关系也类似.然而，不同阶SPP对θ的依赖关系却有着较大区别.图3以Er为例展示了场量在SPP波导横截面中的分布：

The steps of the proposed ADBI are as follows:

(1) If |β|>βT, where βT is a set threshold, then abnormal driving caused by direction offset is determined and a warning signal is output directly. Furthermore, if β>0, it means the vehicle goes toward left; if β<0, it means the vehicle goes toward right. Otherwise, if |β|≤βT, then turn to the step (2), position offset will be used for further judgement.

(2) If |l|>lT, where lT is also a set threshold, then abnormal driving caused by position offset is determined and a warning signal is also output. Furthermore, if l>0, it means the vehicle goes toward right; if l<0, it means the vehicle goes toward left. If |l|≤lT, it means the vehicle is traveling normally.

(1) Three 4×24 windows side by side are named A, B and C. The gray average of the pixels in each window is denoted by mA, mB and mC, respectively. Based on the analysis above, if there is a real lane across region B, such as A1B1C1 in Fig.4, then three gray averages of three windows will satisfy the following inequation[13]:

Assume that the linear equations of two detected lanes are y=kLx+bL and y=kRx+bR, the linear equation of the angular bisector can be obtained:

(18)

Thus, β can be expressed as

(19)

According to the description above, l represents the position offset from the middle of the image to the calculated vanishing point.

l=xm-xM

(20)

where xm is the abscissa of the middle of the image, and xM is the abscissa of the detected vanishing point.

The flow chart of the proposed ADBI is shown in Fig.8.

4　Experimental results

In order to verify the performance of the proposed ADBI method, experiments are conducted in the Nanjing-Zhenjiang expressway, which is 81.7km long. Three men and two women participated in the experiments. Their ages ranged from 22 to 36 years, and they had more than three years of driving experience each. We use an on-board camera to acquire lane images, total 12800 frames of images involving various traffic scenes are used to test the proposed ADBI method.

Some results of lane and vanishing point detection are shown in Fig.9, where Fig.9(a) and Fig.9(b) show the detection on cloudy days, Fig.9(c) and Fig.9(d) show the detection at dusk and Fig.9(e) and Fig.9(f) show the detection with the interference from other vehicles. Fig.9(g) shows the result detected on curved road. Fig.9(h) shows the result detected at night. In Fig.9, the solid black line represents the detected lane and the black dot represents the calculated vanishing point. Seen from the Fig.9, although under severe traffic scenes, such as low illuminance at night and occlusion from surrounding vehicles, the proposed method is still able to correctly detect the real lanes and vanishing points, and the correct rate can reach 97.4%.

Fig.8　The flow chart of ADBI

Fig.9　Experiment of lane and vanishing point detection

Table1 gives out the test results of the proposed ADBI method, where the values of threshold βT and lT are related to camera’s parameters and the fixed position. In this paper, βT=15° and lT=50.

Table 1　Test results of ADBI corresponding to Fig.9

NumberLeftlinekLαL(°)RightlinekRαR(°)DirectionoffsetPositionoffsetβ(°)l(pixel)TestresultFig．9(a)-0．484-25．8212．50068．19921．19⁃⁃leftdeviation(duetodirectionbiasing)Fig．9(b)-0．636-32．4712．37067．12217．33⁃⁃leftdeviation(duetodirectionbiasing)Fig．9(c)-3．039-71．7840．48225．747-23．02⁃⁃rightdeviation(duetodirectionbiasing)Fig．9(d)-2．883-70．8720．72836．050-17．41⁃⁃rightdeviation(duetodirecionbiasing)Fig．9(e)-0．887-41．5762．12264．76711．6027travelingnormallyFig．9(f)-1．228-50．8421．78860．7814．9730travelingnormallyFig．9(g)-0．869-41．0000．73436．294-2．35-45travelingnormallyFig．9(h)-0．692-34．7001．63258．500-1．3661leftdeviation(duetopositionbiasing)

Three typical indicators, such as precision, recall, and accuracy, are used to further evaluate performance of ADBI. Their definitions are as follows[13]: precision=TP/(TP+FP), recall=TP/(TP+FN), and accuracy=(TP+TN)/(TP+FN+FP+TN), where, TP, FP, FN and TN are the abbreviations of true positives, false positives, false negatives, and true negatives, respectively.

坚持正向的评价引导非常重要。教师在教学过程中评价要以个体评价为主，而且评价的内容我认为不应过短，不能每次都是“好”“很好”而应该是变为“表现很棒，再来两句”类似的话术，让学生阅读内容结束时不是终点，而是阅读的新开始。

Experimental results of ADBI under various traffic scenes are shown in Table 2, where “under good scenes” means that the test samples are captured under good illumination or non-occlusion, “under severe scenes” means that the test samples are captured under poor illumination or occlusion from surrounding vehicles.

Compared with state-of-the-art methods[14, 15], the proposed method reaches 97.4% in precision of ADBI.However, the precisions in Refs[14] and [15] are 92% and 90%, respectively.This shows that the proposed method outperforms the state-of-the-art methods.

Table 2　　　　Performance of the proposed ADBI method under various traffic scenes

　Precision(%)Recall(%)Accuracy(%)Undergoodscenes97．494．798．4Underseverescenes95．892．896．3

5　Conclusions

This paper proposes a new ADBI method based on direction and position offsets, where a two-factor identification strategy is used to improve the accuracy and reliability of the ADBI. Self-adaptive edge detection based on Sobel operator is used to extract the edge information of lanes. A Hough transform based on local search scope is employed to quickly detect the lane, and a validation scheme based on priori information is proposed to further verify the detected lane, which enhances the efficiency and reliability of lane detection. Experimental results indicate that the proposed algorithm has strong robustness under complex road conditions.

References

[ 1] Hu J. Comparison of different features and classifiers for driver fatigue detection based on a single EEG channel[J]. Computational & MathematicalMethodsinMedicine, 2017, 2017:5109530

[ 2] Fu R, Wang H, Zhao W, et al. Dynamic driver fatigue detection using hidden Markov model in real driving condition[J]. ExpertSystemsWithApplications, 2016, 63(30): 397-411

[ 3] Jo J, Lee S J, Kang R P, et al. Detecting driver drowsiness using feature-level fusion and user-specific classification[J]. ExpertSystemswithApplications, 2014, 41(4):1139-1152

[ 4] Zhao C, Zhang X, Zhang B, et al. Driver’s fatigue expressions recognition by combined features from pyramid histogram of oriented gradient and contourlet transform with random subspace ensembles[J]. IEEETransactionsonIntelligentTransportationSystems, 2013, 7(1): 36-45

[ 5] Sandström M, Lampsijärvi E, HolmströmA, et al. Detecting lane departures from steering wheel signal[J]. AccidentAnalysisandPrevention, 2017, 99: 272-278

[ 6] McDonald A D, Lee J D, Schwarz C, et al.Steering in a random forest: ensemble learning for detecting drowsiness-related lane departures[J].HumanFactors, 2014, 56(5): 986-998

[ 7] Cheng H Y, Jeng B S, Tseng P T, et al. Lane detection with moving vehicles in the traffics scenes[J]. IEEETransactionsonIntelligentTransportationSystems, 2006,7(4):571-582

[ 8] Feng H, Chen Y, Xu Y. Real-time image acquisition and sobel edge detection based on FPGA[J]. TransducerandMicrosystemTechnologies, 2011, 30(6): 116-118

[ 9] Yoo H, Yang U, Sohn K. Gradient-enhancing conversion for illumination-robust lane detection[J]. IEEETransactionsonIntelligentTransportationSystems, 2013, 14(3):1083-1094

[10] Yang X, Duan J, Gao D, et al. Research on lane detection based on improved Hough transform[J]. ComputerMeasurement &Control, 2010, 18(2): 292-294, 298

[11] Satzoda R, Sathyanarayana S, Srikanthan T. Hierarchical additive Hough transform for lane detection[J]. IEEETransactionsonEmbeddedSystemsLetters, 2010, 2(2): 23-26

[12] Li Q, Wang F, Hu X, et al. A fast lane detection algorithm based on brightness difference[C]. In: Proceedings of the 11th International Computer Conference on Wavelet Active Media Technology and Information Processing, Chengdu, China, 2014. 253-256

[13] Powers D M W. Evaluation: From precision, recall and f-factor to ROC, informedness, markedness &correlation[J]. JournalofMachineLearningTechnologies, 2011, 2:2229-3981

[14] Xu L, Hu S, Luo Q. A new lane departure warning algorithm consideringthe driver’s behavior characteristics[J]. MathematicalProblemsinEngineering, 2015, 2015: 1-11

[15] Yalic H Y, Keceli A S, Kaya A. On-board driver assistance system for lane departure warning and vehicle detection[J]. InternationalJournalofElectricalEnergy, 2013, 1(3):132-136