【摘要】：近些年來,基于計算機視覺的行人檢測與跟蹤技術(shù)得到了突飛猛進的發(fā)展,已有大量的應(yīng)用出現(xiàn)在視頻場景監(jiān)控,目標行為分析,機器人控制,人機接口,智能交通等領(lǐng)域。一些比較熱門而且尚處于發(fā)展階段的應(yīng)用包括汽車輔助駕駛、自主車系統(tǒng)、無人駕駛汽車等。由于面對越來越復(fù)雜的外部應(yīng)用環(huán)境,特別是對于快速變化的環(huán)境背景,純粹基于計算機視覺的目標識別技術(shù)正面臨越來越大的挑戰(zhàn)。就目前而言,依然缺乏一種魯棒、精確并且快速的檢測與跟蹤算法。以行人檢測為例,行人存在非常大的類內(nèi)變化,特別是衣著、光照的顯著變化,加上人體姿態(tài)和運動的隨意性,以及人與人與環(huán)境之間的相互影響,和類人物體的干擾,使得當前各類行人檢測算子,離復(fù)雜環(huán)境下的實際需求均還有一定差距。而對于應(yīng)用范圍更廣的目標跟蹤技術(shù),靜態(tài)背景下的相關(guān)研究已經(jīng)比較成熟,而對于移動變化的背景,目前的研究還非常不足。本文的主要貢獻包括以下幾個方面:1)研究了單幀圖像的快速行人檢測算法。隨著基于視覺的行人檢測算法的研究深入,進一步提高行人檢測算子的性能已面臨非常大的困難。同時,復(fù)雜的算法帶來了更高的計算復(fù)雜度,嚴重影響了檢測系統(tǒng)的時效性。我們考慮在單幀圖像條件下,通過適當?shù)母倪M使算法在不降低時效性的情況下,進一步提高其魯棒性。主要研究了基于Adaboost+Chn Ftrs的快速行人檢測算法,通過選擇不同的特征組合得到了相對最優(yōu)的算子。此外,還在Adaboost的學習過程中采用了特征查找表LUT算法,大大提高了訓練速度,并提高了訓練數(shù)據(jù)的樣本容量。2)對不同在線分類算法與目標特征模型進行了跟蹤性能的測試,并提出基于Fern分類算法的在線目標顏色模型,和結(jié)構(gòu)模型,經(jīng)過測試取得了更好的跟蹤效果。針對常用的特征,如顏色特征、結(jié)構(gòu)特征、模板等,結(jié)合不同的在線分類算法建立在線目標分類模型,其中顏色特征采用基于超像素顏色直方圖的SPT算法;結(jié)構(gòu)特征采用基于壓縮特征的CT算法,對于CT,本文也提出采用基于多通道圖像的壓縮特征,可改善算法效果;模板匹配采用最近鄰分類器(NNClassifier)和TLD算法,通過測試指出了各種目標特征在線分類模型的適應(yīng)場景和各自的不足。在此基礎(chǔ)上,本文提出了采用Fern算法結(jié)合顏色超像素特征,和多通道空間小塊特征,實現(xiàn)了更好的跟蹤效果,可以較好地適應(yīng)包括存在各種遮擋、光照變化、目標姿態(tài)變化、目標尺度變化,和背景多樣變化的復(fù)雜場景。3)考慮目標跟蹤的一般情況,為了更好地適應(yīng)目標尺度的變化,提出采用基于粒子群優(yōu)化的粒子濾波作為跟蹤濾波算法,實驗證明,粒子群能顯著提高粒子濾波的濾波性能,并對跟蹤整體性能的提高有很大作用。4)提出了一種移動背景和復(fù)雜場景下的針對行人等目標的相對實時魯棒的跟蹤算法,可處理包括目標姿態(tài)、尺度變化,光照變化,遮擋和干擾等各種跟蹤場景。算法采用多元特征在線目標分類模型,結(jié)合基于粒子群優(yōu)化的粒子濾波算法進行跟蹤。其多元特征模型包含F(xiàn)ern顏色模型、Fern結(jié)構(gòu)模型、以及CT自適應(yīng)結(jié)構(gòu)模型,各特征模型之間具有很強的互補性,綜合后可達到相當不錯的跟蹤效果。此外,在該算法中引入基于直接模板匹配的NNClassifier作為監(jiān)督模型,在不增加計算復(fù)雜度的情況下利用模板的慢自適應(yīng)性抵制跟蹤漂移。我們還提出了如何解決光照變化問題的方法(其中顏色模型對光照變化特別敏感),使得算法可以處理復(fù)雜光照變化條件下的目標跟蹤問題。
[Abstract]:In recent years, the technology of pedestrian detection and tracking based on computer vision has been developed by leaps and bounds, and a large number of applications appear in the fields of video scene monitoring, target behavior analysis, robot control, man-machine interface, intelligent traffic and the like. Some of the most popular and developing applications include auto-assisted driving, autonomous vehicle systems, driverless cars, and more. Because of the increasingly complex external application environment, especially for rapidly changing environment, the target recognition technology based on computer vision is facing more and more challenges. At present, a robust, accurate and fast detection and tracking algorithm is still lacking. By taking the pedestrian detection as an example, the pedestrian has a very large intra-class change, in particular the remarkable change of the clothes and the illumination, the random nature of the human body posture and the movement, the mutual influence between the human and the environment, and the interference of the human-like object, so that the present various pedestrian detection operators, There is still a gap in the actual demand in a complex environment. For the target tracking technology with a wider application range, the related research in the static background has become more mature, and for the background of the moving change, the current research is still very low. The main contribution of this paper includes the following aspects:1) The fast pedestrian detection algorithm for single-frame image is studied. With the study of the visual-based pedestrian detection algorithm, it is very difficult to further improve the performance of the pedestrian detection operator. At the same time, the complicated algorithm brings higher computational complexity, which seriously affects the time-effectiveness of the detection system. We consider that, under the condition of single-frame image, the robustness of the algorithm is further improved by appropriate improvement so that the algorithm can not reduce the time-effectiveness. The fast pedestrian detection algorithm based on Adaboost + Chn Ftrs is mainly studied, and the relative optimal operator is obtained by selecting different feature combinations. In addition, the characteristic look-up table LUT algorithm is adopted in the learning process of the Adaboost, the training speed is greatly improved, and the sample capacity of the training data is improved. In addition, the online target color model and the structure model based on the Fern classification algorithm are put forward, and a better tracking effect is obtained through the test. Aiming at the common characteristics, such as color characteristics, structural features, templates and the like, an on-line target classification model is established in combination with different on-line classification algorithms, wherein the color characteristics adopt the SPT algorithm based on the hyper-pixel color histogram; the structure characteristic adopts a CT algorithm based on the compression feature, In this paper, a multi-channel image-based compression feature is proposed, which can improve the algorithm effect; the template matching adopts the nearest neighbor classifier (NNNN) and the TLD algorithm, and the adaptive scene and the respective deficiency of the on-line classification model of various target features are pointed out through the test. On the basis of this, this paper proposes that the Fn algorithm is used to combine the color super-pixel features and the multi-channel space tile features, and the better tracking effect can be realized, which can be better adapted to include various occlusion, light change, target attitude change, target scale change, and in order to better adapt to the change of the target scale, a particle filter based on the particle swarm optimization is adopted as a tracking filtering algorithm, And a relative real-time robust tracking algorithm for pedestrian and the like in a moving background and a complex scene is proposed, and various tracking scenes including target pose, scale change, illumination change, occlusion and interference can be processed. In this paper, a multi-element on-line target classification model is used to track the particle swarm optimization based on particle swarm optimization. The multivariate characteristic model contains the Fern color model, the Fern structure model, and the CT self-adaptive structure model. In addition, the NNNClassifier based on the direct template matching is introduced as the monitoring model in the algorithm, and the tracking drift is resisted by the slow self-adaptability of the template without increasing the computational complexity. We also put forward a method to solve the problem of illumination change (in which the color model is particularly sensitive to light change), so that the algorithm can deal with the problem of target tracking under the condition of complex illumination.
【學位授予單位】：中國科學院研究生院(上海技術(shù)物理研究所)
【學位級別】：博士
【學位授予年份】：2015
【分類號】：TP391.41

【參考文獻】

相關(guān)期刊論文 前5條

1 范保杰;朱琳琳;崔書平;李向軍;唐延東;;旋翼無人機視覺跟蹤系統(tǒng)[J];紅外與激光工程;2011年01期

2 劉足華;熊惠霖;;基于隨機森林的目標檢測與定位[J];計算機工程;2012年13期

3 錢慧佳;楊明;李榚;王春香;;一種基于激光雷達和視覺的行人檢測方法[J];上海交通大學學報;2010年07期

4 李兵;劉磊;魏志強;;一種具有強實時性、強魯棒性的圖像匹配算法[J];軟件學報;2014年07期

5 李娟;邵春福;楊勵雅;李琦;;基于Kalman濾波的行人跟蹤方法研究[J];交通運輸系統(tǒng)工程與信息;2009年06期

，

本文編號：2479766