基于結(jié)構(gòu)支持向量機的目標跟蹤算法研究
發(fā)布時間：2018-10-17 19:09
【摘要】：隨著科技進步和社會發(fā)展,計算機視覺跟隨人工智能的腳步走入人類視野。目標檢測與跟蹤課題,作為計算機視覺的關(guān)鍵問題,也是一個經(jīng)典難題,近年來受到各個相關(guān)領(lǐng)域研究學者的關(guān)注,并且應(yīng)對不同場景探索不同的檢測與跟蹤算法。在目標檢測和目標跟蹤兩個領(lǐng)域中,關(guān)鍵的問題都在于如何有效描述目標、如何讓計算機準確識別目標;不同點在于檢測看重的是精確度,而跟蹤在于實時性。應(yīng)對這兩種需求,本文結(jié)合支持向量機的優(yōu)秀分類特性,研究了以下檢測和跟蹤系統(tǒng)。對于目標檢測系統(tǒng),在其訓練階段,首先在每個滑動窗口中分別計算HOG特征與LBPHF特征,然后將兩者結(jié)合構(gòu)成聯(lián)合特征。接著利用線性支持向量機(SVM)訓練分類器,其中本算法通過自舉法(Bootstrap Method)不斷更新優(yōu)化分類器,以此獲得最優(yōu)判別模型。在訓練階段的基礎(chǔ)上,將提取所得的聯(lián)合特征輸入上一階段所獲得的分類器中進行判別,最后采用非極大值抑制(NMS)的融合方法對重疊檢測窗口進行融合,以此獲得最終的檢測結(jié)果。實驗證明改進后的方法滿足檢出率高、計算復雜度低、抗行人肢體偏轉(zhuǎn)干擾能力強等要求。對于目標跟蹤系統(tǒng),首先利用無模型的跟蹤框架,運用改進的HOG-LBPHF對目標進行表觀,并且結(jié)合目標間的結(jié)構(gòu)信息,以此來訓練SVM。其次采用被動主動感知器對分類平面進行優(yōu)化。最后用最小生成樹模型確定下一幀的所在位置。經(jīng)過實驗對比,本算法具有良好的跟蹤性能。
[Abstract]:With the progress of science and technology and social development, computer vision follows the pace of artificial intelligence into human vision. Target detection and tracking, as a key problem of computer vision, is also a classical problem. In recent years, it has attracted the attention of researchers in various related fields, and different detection and tracking algorithms should be explored in different scenes. In the two fields of target detection and target tracking, the key problems are how to describe the target effectively and how to accurately identify the target by computer. In response to these two requirements, this paper studies the following detection and tracking systems combined with the excellent classification characteristics of support vector machines. In the training phase of the target detection system, the HOG feature and the LBPHF feature are calculated in each sliding window, and then the two features are combined to form a joint feature. Then the classifier is trained by linear support vector machine (SVM), where the optimal discriminant model is obtained by updating the optimal classifier by bootstrap (Bootstrap Method). On the basis of the training stage, the extracted joint features are input into the classifier obtained in the previous stage to discriminate. Finally, the overlapping detection window is fused using the fusion method of non-maximum suppression (NMS). Finally, the final test results are obtained. The experimental results show that the improved method meets the requirements of high detection rate, low computational complexity and strong anti-pedestrian limb deflection interference. For the target tracking system, we first use the model-free tracking framework, use the improved HOG-LBPHF to visualize the target, and combine the structure information between the targets to train the SVM.. Secondly, passive active perceptron is used to optimize the classification plane. Finally, the location of the next frame is determined by the minimum spanning tree model. The experimental results show that the algorithm has good tracking performance.
【學位授予單位】：哈爾濱理工大學
【學位級別】：碩士
【學位授予年份】：2016
【分類號】：TP391.41;TP18
本文編號：2277675
[Abstract]:With the progress of science and technology and social development, computer vision follows the pace of artificial intelligence into human vision. Target detection and tracking, as a key problem of computer vision, is also a classical problem. In recent years, it has attracted the attention of researchers in various related fields, and different detection and tracking algorithms should be explored in different scenes. In the two fields of target detection and target tracking, the key problems are how to describe the target effectively and how to accurately identify the target by computer. In response to these two requirements, this paper studies the following detection and tracking systems combined with the excellent classification characteristics of support vector machines. In the training phase of the target detection system, the HOG feature and the LBPHF feature are calculated in each sliding window, and then the two features are combined to form a joint feature. Then the classifier is trained by linear support vector machine (SVM), where the optimal discriminant model is obtained by updating the optimal classifier by bootstrap (Bootstrap Method). On the basis of the training stage, the extracted joint features are input into the classifier obtained in the previous stage to discriminate. Finally, the overlapping detection window is fused using the fusion method of non-maximum suppression (NMS). Finally, the final test results are obtained. The experimental results show that the improved method meets the requirements of high detection rate, low computational complexity and strong anti-pedestrian limb deflection interference. For the target tracking system, we first use the model-free tracking framework, use the improved HOG-LBPHF to visualize the target, and combine the structure information between the targets to train the SVM.. Secondly, passive active perceptron is used to optimize the classification plane. Finally, the location of the next frame is determined by the minimum spanning tree model. The experimental results show that the algorithm has good tracking performance.
【學位授予單位】：哈爾濱理工大學
【學位級別】：碩士
【學位授予年份】：2016
【分類號】：TP391.41;TP18
【參考文獻】
相關(guān)期刊論文 前5條
1 黃凱奇;陳曉棠;康運鋒;譚鐵牛;;智能視頻監(jiān)控技術(shù)綜述[J];計算機學報;2015年06期
2 孫凱;嚴瀟然;謝榮平;;基于手勢識別的智能家居人機交互系統(tǒng)設(shè)計[J];工業(yè)控制計算機;2014年04期
3 張春鳳;宋加濤;王萬良;;行人檢測技術(shù)研究綜述[J];電視技術(shù);2014年03期
4 楊利平;辜小花;;用于人臉識別的相對梯度直方圖特征描述[J];光學精密工程;2014年01期
5 孫銳;陳軍;高雋;;基于顯著性檢測與HOG-NMF特征的快速行人檢測方法[J];電子與信息學報;2013年08期
，本文編號：2277675
本文鏈接：http://sikaile.net/kejilunwen/zidonghuakongzhilunwen/2277675.html
最近更新
教材專著
