05中英文雙語畢設外文翻譯成品v2v協(xié)同環(huán)境中對智能車輛進行快速行人檢測和動態(tài)跟蹤

1、<p>　　外文標題：Fast pedestrian detection and dynamic tracking for intelligent vehicles within V2V cooperative environment</p><p>　　外文作者：Fuliang Li, Ronghui Zhang , Feng You</p><p>　　文獻出處：《Iet I

2、mage Processing》 , 2017 , 11 (10) :833-840</p><p>　　英文2203單詞， 14998字符，中文3668漢字。</p><p>　　此文檔是外文翻譯成品，無需調整復雜的格式哦！下載之后直接可用，方便快捷！只需二十多元。</p><p><b>　　原文：</b></p><p

3、>　　Fast pedestrian detection and dynamic tracking for intelligent vehicles within V2V cooperative environment</p><p>　　Fuliang Li, Ronghui Zhang , Feng You</p><p>　　Abstract: </p><

4、p>　　Pedestrian detection has become one of the hottest topics in intelligent traffic system because of its potential applications in driver assistance and automatic driving. In this study, a fast pedestrian detection

5、and dynamic tracking method within vehicle-to-vehicle (V2V) cooperative environment is proposed. A dynamic tracking-by-detection framework for real-time pedestrian detection is developed. First, a cascade classifiers, ba

6、sed on selected Haar-like features, is trained to detect pedestrian</p><p>　　Introduction</p><p>　　In recent years, pedestrian deaths resulting from the complex traffic environment accounted for

7、 60% of all deaths on the roads [1]. Aiming to reduce collision and danger to pedestrians from traffic, pedestrian active safety analysis has become an international research focus, especially pedestrian detection techno

8、logy.</p><p>　　In general, pedestrian detection methods can be divided into target characteristics template-based and pedestrian-based learning methods. The former type of methods cost less and are relativel

9、y simple. However, those methods only work well by detected obvious contour, and their detection effects have a direct relationship with template choice. Davis and Mark [2] proposed a two-step template method based on in

10、frared images. Detection results are correlated with the selected template directly. A p</p><p>　　Pedestrian tracking is expected to predict information such as pedestrian's position in the next few fram

11、es based on the detection information in the current frame. In general, continuous detection can be replaced by pedestrian tracking for enhancing the pedestrian detection's real-time performance [14]. Probability-bas

12、ed pedestrian tracking method is a research hotspot for solving tracking problems. Without loss of generality, pedestrian tracking can be treated as a state estimation issue. The </p><p>　　model based probab

13、ility distribution to estimate pedestrian maximum likelihood moving direction directly.</p><p>　　The key contributions of this proposed method include: a dynamic tracking-by-detection method for real-time pe

14、destrian detection, which means that selected Haar-like features based cascade classifiers are proposed to detect pedestrian first, and then make a dynamic pedestrian tracking using CamShift algorithm combined with exten

15、ded Kalman filtering (EKF). A smart phone based V2V cooperative warning system is developed to share useful detection results within blind spots.</p><p>　　Selected Haar-like based cascade classifiers for fas

16、t pedestrian detection</p><p>　　Selected Haar-like features and weak classifier training</p><p>　　Haar-like features are defined as the differences in the greyscale sum of black and white rectan

17、gles’ corresponding regions in the image sub-window, which can extract image texture features effectively. For pedestrian detection, the most frequent changes of greyscale are in the vertical and horizontal directions. T

18、hus, this paper selects global eight types or local human shape rectangular feature shown in Fig. 1, from which the differences of characteristics in appearance can be highlighted effe</p><p>　　Haar-like fea

19、tures can be computed rapidly using intermediate representation called the integral image [22]. The integral image at location (x, y), denoted by ii(x, y), contains the sum of the pixel values above and to the left of (x

20、, y) shown in Fig. 2a, which can</p><p>　　be calculated by</p><p>　　where I(x′, y′) is the original image that can be obtained using following iterative calculations:</p><p>　　Thus,

21、 the integral image can be computed in just one pass over the original image, which means that we can calculate the value of selected Haar-like features rapidly. Take four array references shown in Fig. 2b as an example,

22、 the integral image value at location 1 for the regional grey level A sum is denoted by A. Correspondingly, the values of locations 2, 3, and 4 are A + B, A + C, and A + B + C + D. Finally, the sum within rectangular reg

23、ion D can be expressed as 4 + 1?(2 + 3).</p><p>　　During the classifiers training process, an improved Adaboost algorithm was used to process weak classifiers training based on minimum error rate principle,

24、which can reduce the weight of samples with the correct classification, but increase the weight of samples with the wrong classification. More precisely, the weak offline trained classifiers can decrease the error rate a

25、nd shorten the training time at the same time. Then, strong classifiers can be obtained, which are composed of several weak</p><p>　　Design of the cascade classifier</p><p>　　Cascade structure,

26、a kind of degeneration decision tree, focuses on processing the key image region [23]. Each strong classifier search window is moved across the input target and check whether there is pedestrian or not. Only the input ta

27、rget passing through all strong classifiers can be considered as pedestrian. The flowchart of cascade classier is shown in Fig. 3.</p><p>　　V2V cooperative warning platform</p><p>　　There are so

28、me typical blind spots in urban traffic conditions, such as turning, lane-changing areas at intersection or merging sections, where the above pedestrian detection and tracking methods cannot work effectively. In this cas

29、e, using V2V communication is one of ways to share pedestrian warning information out of individual vehicle detection range. Based on those timely cooperative warning information, the driver can make more reliable decisi

30、ons and has a better chance of reacting properly</p><p>　　To reduce pedestrian collision probability, we propose a novel cooperative warning framework to share pedestrian warning information within V2V coope

31、rative environment. Detected pedestrian warning information, including pedestrian's position, speed, and movement, can be shared to related vehicles in blind spots. The above position interruption and redundant infor

32、mation are two significant issues which this framework has tried to deal with. To reduce the impact of position interruption, the frame</p><p>　　More precisely, the application scenario description is as fol

33、lows: a host vehicle detects and tracks pedestrians using the above proposed fast pedestrian detection and dynamic tracking method, which was introduced in Sections 2 and 3. Then its GPS data and pedestrian GPS data esti

34、mated by single-frame static image distance model are sent to the BIPC via 4G network [28]. After that, the BIPC determines the pedestrian speed and moving direction through analysing pedestrian GPS data sequences firs&l

35、t;/p><p>　　Experiments</p><p>　　To validate the proposed method, we develop a system using Visual 2010 and Intel OpenCV to test it. Then, it is transplanted to the HUAWEI GlORY mobile phone. It has

36、 hardware configurations of Hisilicon Kirin 935 + 3GB RAM + 2000W BSI camera to realise fast pedestrian detection. All test equipments were installed into a Chery Tiggo NCV vehicle shown in Fig. 7.</p><p>　　

37、Conclusion</p><p>　　In this paper, we present a fast pedestrian detection and dynamic tracking method for intelligent vehicles within V2V cooperative environment. The key contributions of this proposed metho

38、d include: a dynamic tracking-by-detection method for real-time pedestrian detection, which means that selected Haar-like features based cascade classifiers are proposed to detect pedestrian first. Then in view of the no

39、n-linear characteristics of urban road conditions, a dynamic pedestrian tracking using CamShi</p><p>　　However, this paper cannot analyse the effect of bad weather (such as rain, fog day) and vehicle speed o

40、n the algorithm's performance and reliability [32] due to experiment environment lmitation. All of those will be the focus of our future research work.</p><p>　　References</p><p>　　Zhang, S.

41、, Christian, B., Armin, B.C.: ‘Efficient pedestrian detection via rectangular features based on a statistical shape model’, IEEE Trans. Intell. Transp. Syst., 2015, 16, (2), pp. 763-775</p><p>　　Davis, J.W.,

42、 Mark, A.K.: ‘A two-stage template approach to person detection in thermal imagery’, IEEE Workshop Motion Video Comput., 2005, 2005, pp. 364-369</p><p>　　Bertozzi, M., Broggi, A., Del, R.M., et al.: ‘A pedes

43、trian detector using histograms of oriented gradients and a support vector machine classifier’. IEEE Conf. on Intelligent Transportation Systems, 2007, pp. 143-148</p><p>　　Zhuang, J., Liu, Q.: ‘Nighttime pe

44、destrian detection method for driver assistance systems’, J. South China Univ. Technol., 2012, 40, (8), pp. 56-62</p><p>　　Li, F., You, F., Zhang, R., et al.: ‘An improved real-time detection and localizatio

45、n scheme for pedestrian based on information fusion’, Int. J. Appl. Math. Stat., 2013, 51, (22), pp. 99-107</p><p>　　Can, Y, Li, B., Xu, G.: ‘Particle filter based multi-pedestrian tracking by HOG and HOF’.

46、4th IEEE Int. Conf. on Information Science and Technology, 2014, pp. 714-717</p><p>　　Guo, L., Zhang, M., Li, L., et al.: ‘Body parts features based pedestrian detection for active pedestrian protection syst

47、em’, Promet — Traffic — Traffico, 2016, 28, (2), pp. 113-142</p><p>　　Yao, S., Pan, S., Wang, T., et al.: ‘A new pedestrian detection method based on combined HOG and LSS features’, Neurocomputing, 2015, 151

48、, (2015), pp. 1006-1014</p><p>　　Dollar, P, Wojek, C., Schiele, B., et al.: ‘Pedestrian detection: an evaluation of the state of the art’, IEEE Trans. Pattern Anal. Mach Intell., 2012, 34, (4), pp. 743-761&l

49、t;/p><p>　　Oliveira, L., Urbano, N., Paulo, P.: ‘On exploration of classifier ensemble synergism in pedestrian detection’, IEEE Trans. Intell. Transp. Syst., 2010, 11,</p><p>　　,pp. 16-27</p>

50、;<p>　　Ge, J., Luo, Y., Tei, G.: ‘Real-time pedestrian detection and tracking at nighttime for driver-assistance systems’, IEEE Trans. Intell. Transp. Syst., 2009, 10, (2), pp. 283-298</p><p>　　Xu, Y.

51、, Xu, D., Lin, S., et al.: ‘Detection of sudden pedestrian crossings for driving assistance systems’, IEEE Trans. Syst. Man Cybern B, Cybern., 2012, 42, (3), pp. 729-739</p><p>　　Sun, H., Cheng, W., Wang, B.

52、, et al.: ‘Pyramid binary pattern features for real-time pedestrian detection from infrared videos’, Neurocomputing, 2011, 74, (5), pp. 797-804</p><p>　　Dollar, P., Wojek, C., Schiele, B., et al.: ‘Pedestria

53、n detection: a benchmark’. IEEE Conf. on Computer Vision and Pattern Recognition, 2009, pp. 304~311</p><p>　　Levi, D., Silberstein, S., Bar-Hillel, A.: ‘Fast multiple-part based object detection using kd-fer

54、ns’. IEEE Conf. on Computer Vision and Pattern Recognition, 2013, pp. 947-954</p><p><b>　　譯文：</b></p><p>　　V2V協(xié)同環(huán)境中對智能車輛進行快速行人檢測和動態(tài)跟蹤</p><p>　　Fuliang Li, Ronghui Zhang

55、, Feng You</p><p><b>　　摘要</b></p><p>　　在智能交通系統(tǒng)中，行人檢測已經(jīng)成為最受熱議的話題之一，這是因為它在駕駛輔助和自動駕駛中具有潛在的應用。在本次的研究中，我們提出了車輛間（V2V）協(xié)同環(huán)境中的快速行人檢測和動態(tài)跟蹤方法。本文開發(fā)了一種用于實時行人檢測的動態(tài)跟蹤檢測框架。首先，對具有選定的Haar-like特征的級聯(lián)分類器

56、進行訓練以檢測行人。 然后，結合CamShift算法與拓展的卡爾曼濾波來用于行人動態(tài)跟蹤。 最后，利用眾包檢測信息，開發(fā)了基于智能手機的V2V合作預警系統(tǒng)，其可以以在盲區(qū)內共享有用的檢測結果。 實驗結果表明，該方法具有實時性和準確性，可為道路交通安全監(jiān)測技術提供參考。</p><p><b>　　引言</b></p><p>　　近年來，由于復雜的交通環(huán)境造成的行人死

57、亡占道路死亡人數(shù)的60％[1]。 為了減少交通行人碰撞以及危險，交通路面行人安全分析已成為國際研究熱點，特別是行人檢測技術。</p><p>　　一般來說，行人檢測方法可以分為基于目標特征和基于了解行人行為的方法。前一種方法成本較低，而且相對簡單。然而，這些方法只能檢測到一般的輪廓，并且其檢測效果與模板選擇有直接關系。 Davis和Mark [2]提出了基于紅外圖像的兩步模板方法。檢測結果直接與選定的模板相關聯(lián)。

58、 Bertozzi等人提出了一種基于步行模式的檢測方法[3]。該方法首先基于行人步態(tài)模式計算人體概率模板，然后使用計算的聯(lián)合概率確定物體是否為行人。該方法適用于檢測可見腿部的行人。 Zhuang和Liu [4]提出了一種概率模板匹配算法來實現(xiàn)行人檢測。該方法使用局部雙分割閾值來提取候選目標并運用多尺度概率模板。這種方法需要的樣本較少，但在復雜的城市道路環(huán)境中誤差率較高，實時性較差。</p><p>　　在當前的框

59、架中基于監(jiān)測的信息，行人跟蹤系統(tǒng)將被用于預測諸如行人在接下來幾幀的位置等信息。一般來說，連續(xù)性檢測可以被行人追蹤系統(tǒng)所取代，以提高行人檢測的實時性[14]。基于概率的行人跟蹤方法是解決跟蹤問題的研究熱點。在不失其普遍性的情況下，行人追蹤可被視為行人狀態(tài)預測的問題?？柭鼮V波和粒子濾波跟蹤方法是該領域的常用方法[15]。劉等人[16]提出了CamShift運動目標跟蹤算法，使用擴展的卡爾曼濾波來估計目標的運動速度和空間位置。 Li等人[1

60、7]開發(fā)了一種自適應卡爾曼濾波跟蹤算法，實時修改濾波器的統(tǒng)計模型，并應用最小二乘支持向量機來估計目標移動方向。 王和唐[18]提出了一種采用分段高斯分段模型的粒子濾波行人跟蹤方法，這是基于模型的概率分布直接估計行人最大似然移動方向。</p><p>　　圖一 選定的Haar-like特征</p><p><b>　　圖二 積分圖計算</b></p>

61、<p>　　提出的這種方法主要意義包括：它是一種實時行人檢測的動態(tài)跟蹤檢測方法，這意味著具有選定的Haar-like特征的級聯(lián)分類器被首次提出來檢測行人，然后使用動態(tài)行人跟蹤 CamShift算法并結合擴展的卡爾曼濾波（EKF）， 開發(fā)了基于智能手機的V2V協(xié)作預警系統(tǒng)，以在盲點內共享有用的檢測結果。</p><p>　　具有選定的Haar-like特征的級聯(lián)分類器用于快速行人檢測</p>

62、<p>　　選定的Haar-like特征以及弱分類器的訓練</p><p>　　Haar-like特征可以定義為圖像子窗口中黑白矩形對應區(qū)域灰度等級的差異，其可以有效提取圖像紋理特征。 對于行人檢測，最常見的灰度變化是垂直和水平方向。 因此，本文選取了圖1所示的全局八種類型或局部人體形狀矩形特征，從中可以有效地突出表現(xiàn)出特征差異。</p><p>　　可以使用被稱之為積分圖像

63、的中間表達式快速計算類Haar-like特征[22]。積分圖像位置為（x，y），可以表達為ii(x, y),包含在圖2a中示出的（x，y）的上方和左側的像素值的總和，其可以計算為：</p><p>　　其中I（x'，y'）是可以使用以下迭代計算獲得的原始圖像：</p><p>　　因此，積分圖像可以在原始圖像上僅僅計算一次，這意味著我們可以快速計算選擇的Haar-like特

64、征值。 以圖2b中所示的四個陣列參考為例，區(qū)域灰度A sum的位置1處的積分圖像值由A表示。相應地，位置2,3和4的值是A + B，A + C和A + B + C + D。最后，矩形區(qū)域D內的和可以表示為4 + 1-（2 + 3）。</p><p>　　在分類器訓練過程中，基于最小誤差率原理，可以采用改進的Adaboost算法來處理弱分類器訓練，這可以減少正確分類樣本的權重，但也會增加分類錯誤的樣本權重。 更準確

65、地說，弱離線訓練分類器可以同時降低錯誤率并縮短訓練時間。 然后，可以獲得強分類器，它是由幾個使用線性疊加的弱分類器組成的。 關于訓練過程的更多細節(jié)可以參考我們以前的研究工作[5]。</p><p>　　在分類器訓練過程中，采用改進的Adaboost算法處理基于最小誤差率原理的弱分類器訓練，可以減少正確分類樣本的權重，但增加分類錯誤的樣本權重。 更準確地說，弱離線訓練分類器可以同時降低錯誤率并縮短訓練時間。 然后，

66、可以獲得強分類器，它是由幾個使用線性疊加的弱分類器組成的。 關于培訓過程的更多細節(jié)可以參考我們以前的研究工作[5]。</p><p><b>　　級聯(lián)分類器的設計</b></p><p>　　級聯(lián)結構是一種退化決策樹的算法，重點是處理關鍵圖像區(qū)域[23]。 每個強分類器搜索窗口在輸入的目標上移動并檢查是否存在行人。 只有通過所有強分類器的輸入的目標才能被視為行人。 級

67、聯(lián)分類器的流程圖如圖3所示。</p><p>　　圖三 級聯(lián)分類器的設計</p><p>　　V2V 協(xié)同預警平臺</p><p>　　在城市交通道路中存在一些常見的盲點，如轉彎、交叉口車道變換區(qū)或合并路段等，上述的行人檢測跟蹤方法對這些盲點無法有效發(fā)揮作用。在這種情況下，使用V2V通信是在各個車輛超出其檢測范圍之外去共享行人預警信息最好的方法之一。根據(jù)那些及時的協(xié)

68、同預警信息，駕駛員可以做出更可靠的決策，并有更好的應對緊急情況的機會。但是，在行人協(xié)同預警過程中，有幾個挑戰(zhàn)。首先是位置信息的中斷，即全球定位系統(tǒng)（GPS）接收機定位的位置與實際位置之間的差距。不準確的傳感器信息導致不確定的車輛或行人狀態(tài)信息，這會影響協(xié)同預警系統(tǒng)的性能。另一個挑戰(zhàn)是冗余的信息，這意味著可能會使驅動程序承載過多的預警信息。</p><p>　　為了減少行人碰撞概率，我們提出了一個新的協(xié)作預警框架，

69、在V2V協(xié)作環(huán)境中共享行人預警信息。檢測到的行人預警信息（包括行人的位置、速度和移動）可以在相關車輛的盲點中共享。上述的位置中斷和冗余信息是該框架試圖要解決的兩個重要問題。為了降低位置中斷帶來的影響，該框架開發(fā)了一種采用載波相位差分技術的雙GPS定位系統(tǒng)，一方面提高定位精度，另一方面對廣播間隔進行優(yōu)化[27]。更多細節(jié)將在第5節(jié)中討論。在成功的預警與風險干擾的驅動程序之間提出一種權衡機制來釋放冗余信息。在不失一般性的情況下，我們假設車輛

70、行駛路線信息是已知的。因此，權衡問題可以轉化為集群優(yōu)化問題。后臺信息處理中心（BIPC）首先根據(jù)車輛GPS和行人檢測和跟蹤信息，將所有盲點車輛分類為潛在的碰撞車輛和碰撞車輛。然后，BIPC根據(jù)碰撞概率為潛在車輛提供不同等級的預警信息。所提出的框架的流程圖如圖6所示。</p><p>　　更確切地說，其應用場景描述如下：主車輛使用上面在第2節(jié)和第3節(jié)中介紹的快速行人檢測和動態(tài)跟蹤方法來檢測和跟蹤行人。然后，其GPS

71、數(shù)據(jù)和行人GPS數(shù)據(jù)通過單 幀靜態(tài)圖像距離模型通過4G網(wǎng)絡發(fā)送到BIPC [28]。 之后，BIPC首先通過分析行人GPS數(shù)據(jù)序列來確定行人速度和行進方向。 然后，BIPC將一群周圍的車輛置于盲點中以形成潛在的碰撞和碰撞車輛。通過潛在的碰撞車輛和行人GPS數(shù)據(jù)來計算更進一步的碰撞概率。 最后，BIPC向潛在碰撞車輛播報不同等級的及時警報信息。2800</p><p>　　為了驗證所提出方法的有效性，我們開發(fā)了一個

72、使用Visual 2010和Intel OpenCV的測試系統(tǒng)。 然后，它被植入到華為榮耀手機。 它具有海思麒麟935 + 3GB RAM + 2000W BSI攝像頭的硬件配置，可以實現(xiàn)快速行人檢測。 所有測試設備都安裝在圖7所示的奇瑞瑞虎NCV車輛上。</p><p><b>　　結論</b></p><p>　　在本文中，我們提出了V2V協(xié)同環(huán)境下智能車輛的快速

73、行人檢測和動態(tài)跟蹤方法。 這種方法的主要價值包括：實時的行人檢測動態(tài)跟蹤檢測方法，這意味著具有Haar-like特征的級聯(lián)分類器被首次提出來檢測行人。 然后針對城市道路情況的非線性特征，采用CamShift算法結合EKF進行動態(tài)行人跟蹤，以提高行人檢測的實時性。 此外，還開發(fā)了基于智能手機的V2V協(xié)同報警系統(tǒng)，以在盲區(qū)內共享有用的檢測結果，從而減少單個車輛的盲點并降低交叉口處的事故率。 道路實驗結果表明，與其他最先進的方法相比，所提出的

74、快速行人檢測方法具有更強大、更高的性能。</p><p>　　然而，由于實驗環(huán)境模擬，本文不能分析惡劣天氣（如雨，霧天）和車速對算法性能和可靠性的影響[32]。 所有這些將成為我們未來研究工作的重點。</p><p><b>　　參考文獻</b></p><p>　　Zhang, S., Christian, B., Armin, B.C.:

75、‘Efficient pedestrian detection via rectangular features based on a statistical shape model’, IEEE Trans. Intell. Transp. Syst., 2015, 16, (2), pp. 763-775</p><p>　　Davis, J.W., Mark, A.K.: ‘A two-stage temp

76、late approach to person detection in thermal imagery’, IEEE Workshop Motion Video Comput., 2005, 2005, pp. 364-369</p><p>　　Bertozzi, M., Broggi, A., Del, R.M., et al.: ‘A pedestrian detector using histogram

77、s of oriented gradients and a support vector machine classifier’. IEEE Conf. on Intelligent Transportation Systems, 2007, pp. 143-148</p><p>　　Zhuang, J., Liu, Q.: ‘Nighttime pedestrian detection method for

78、driver assistance systems’, J. South China Univ. Technol., 2012, 40, (8), pp. 56-62</p><p>　　Li, F., You, F., Zhang, R., et al.: ‘An improved real-time detection and localization scheme for pedestrian based

79、on information fusion’, Int. J. Appl. Math. Stat., 2013, 51, (22), pp. 99-107</p><p>　　Can, Y, Li, B., Xu, G.: ‘Particle filter based multi-pedestrian tracking by HOG and HOF’. 4th IEEE Int. Conf. on Informa

80、tion Science and Technology, 2014, pp. 714-717</p><p>　　Guo, L., Zhang, M., Li, L., et al.: ‘Body parts features based pedestrian detection for active pedestrian protection system’, Promet — Traffic — Traffi

81、co, 2016, 28, (2), pp. 113-142</p><p>　　Yao, S., Pan, S., Wang, T., et al.: ‘A new pedestrian detection method based on combined HOG and LSS features’, Neurocomputing, 2015, 151, (2015), pp. 1006-1014</p&

82、gt;<p>　　Dollar, P, Wojek, C., Schiele, B., et al.: ‘Pedestrian detection: an evaluation of the state of the art’, IEEE Trans. Pattern Anal. Mach Intell., 2012, 34, (4), pp. 743-761</p><p>　　Oliveira,

83、 L., Urbano, N., Paulo, P.: ‘On exploration of classifier ensemble synergism in pedestrian detection’, IEEE Trans. Intell. Transp. Syst., 2010, 11,</p><p>　　,pp. 16-27</p><p>　　Ge, J., Luo, Y.,

84、 Tei, G.: ‘Real-time pedestrian detection and tracking at nighttime for driver-assistance systems’, IEEE Trans. Intell. Transp. Syst., 2009, 10, (2), pp. 283-298</p><p>　　Xu, Y., Xu, D., Lin, S., et al.: ‘De

85、tection of sudden pedestrian crossings for driving assistance systems’, IEEE Trans. Syst. Man Cybern B, Cybern., 2012, 42, (3), pp. 729-739</p><p>　　Sun, H., Cheng, W., Wang, B., et al.: ‘Pyramid binary patt

86、ern features for real-time pedestrian detection from infrared videos’, Neurocomputing, 2011, 74, (5), pp. 797-804</p><p>　　Dollar, P., Wojek, C., Schiele, B., et al.: ‘Pedestrian detection: a benchmark’. IEE

87、E Conf. on Computer Vision and Pattern Recognition, 2009, pp. 304~311</p><p>　　Levi, D., Silberstein, S., Bar-Hillel, A.: ‘Fast multiple-part based object detection using kd-ferns’. IEEE Conf. on Computer Vi