王文強(qiáng) 劉永洛 馬立武
摘 要:針對(duì)傳統(tǒng)均值漂移跟蹤算法由于目標(biāo)框大小不能變化,尤其當(dāng)目標(biāo)尺度大小發(fā)生較大變化或旋轉(zhuǎn)時(shí)容易導(dǎo)致目標(biāo)丟失的問(wèn)題,提出一種聯(lián)合顏色與背景信息的目標(biāo)框自適應(yīng)調(diào)整跟蹤算法。以經(jīng)典均值漂移算法為主體跟蹤框架,構(gòu)建前景目標(biāo)顏色直方圖,以Bhattacharyya距離與迭代次數(shù)作為判斷下一幀目標(biāo)中心位置的條件,每次迭代通過(guò)在當(dāng)前幀目標(biāo)框區(qū)域內(nèi)建立感興趣目標(biāo)與局部背景空間模型,經(jīng)快速傅里葉變換后計(jì)算當(dāng)前幀與下一幀空間模型,得到尺度調(diào)節(jié)因子,作為每一幀跟蹤窗口大小的權(quán)重,進(jìn)而不斷調(diào)整跟蹤窗口尺度大小。通過(guò)自適應(yīng)調(diào)整每一幀跟蹤窗口的尺度調(diào)節(jié)因子,達(dá)到實(shí)時(shí)修正目標(biāo)模型描述,進(jìn)而提高跟蹤準(zhǔn)確性的目的,大大降低了由于目標(biāo)模型固化導(dǎo)致的中心位置跟蹤累積誤差。通過(guò)對(duì)兩組圖像的序列仿真結(jié)果表明,改進(jìn)算法相比于經(jīng)典算法具有更強(qiáng)的魯棒性。
關(guān)鍵詞:均值漂移;自適應(yīng)算法;顏色信息;背景信息;目標(biāo)跟蹤
DOI:10. 11907/rjdk. 181851
中圖分類號(hào):TP312文獻(xiàn)標(biāo)識(shí)碼:A文章編號(hào):1672-7800(2019)002-0038-04
Abstract: For the traditional mean shift tracking algorithm, since the size of the target frame cannot be changed, especially when the size of the target scale changes or rotates, the problem of loss of the target easily happens, so an adaptive adjustment algorithm based on spatial information is proposed. Using the classical mean shift algorithm as the main body tracking framework, the color histogram of the foreground object is constructed. The condition of the target center position of the next frame is determined by the Bhattacharyya distance and the number of iterations. In each iterative procedure, the interest target and the local background space model are established within the current frame. The scale adjustment factor is obtained by calculating the current frame and the next frame after the fast Fourier transform. The adjustment factor is used as the weight of each frame tracking window size and then the tracking window size is constantly adjusted to improve the accuracy of the tracking, which greatly reduces the tracking error of the cumulative center position due to the solidification of the target model. Simulation results of 2 sets of pertinent image sequences show that the improved algorithm is more robust than the classical algorithm.
Key Words: mean shift; self-adaption algorithm; color information; background information; object tracking
0 引言
目標(biāo)跟蹤作為智能交通領(lǐng)域的重要研究課題,對(duì)于保障城市交通順暢建設(shè)智慧城市具有重要意義[1]。運(yùn)動(dòng)目標(biāo)跟蹤實(shí)質(zhì)為在一系列圖像序列中找到感興趣目標(biāo)并標(biāo)注出來(lái),而Mean Shift算法作為經(jīng)典的跟蹤算法之一,是由Fukunaga等[2-5]提出的一種非參數(shù)概率密度梯度估計(jì)方法,其因具有計(jì)算量小、不需要設(shè)置參數(shù)等優(yōu)點(diǎn)得到了廣泛應(yīng)用,而實(shí)時(shí)性好是其最突出的優(yōu)勢(shì)[6],但該算法在迭代過(guò)程中易陷入局部最優(yōu)解。目前衍生出許多改進(jìn)算法,例如通過(guò)對(duì)感興趣目標(biāo)融合角點(diǎn)[7]、紋理[8-10]、梯度[11]、尺度不變特征變換[12]等特征,以減少光照變化及噪聲影響,從而提高跟蹤準(zhǔn)確率,或通過(guò)融合直方圖特性[13]、卡爾曼濾波[14-15]、粒子濾波[16]等經(jīng)典算法解決遮擋、背景相似等問(wèn)題,從而提高算法魯棒性。文獻(xiàn)[17]通過(guò)對(duì)RGB顏色空間各個(gè)分量進(jìn)行加權(quán)建模,進(jìn)而建立可區(qū)分度,可一定程度上提高算法魯棒性,但針對(duì)相似背景干擾的情況效果較差;文獻(xiàn)[18]利用Adaboost弱分類器進(jìn)行特征自適應(yīng)融合,跟蹤精度提高但計(jì)算量較大,因此實(shí)時(shí)性較差。近幾年匹配型跟蹤算法成為研究熱點(diǎn),文獻(xiàn)[19]利用模板空間的稀疏性,將特征從高維降到低維進(jìn)而實(shí)現(xiàn)跟蹤,從而在出現(xiàn)遮擋、光照變化時(shí),具有更強(qiáng)的魯棒性。但其對(duì)前景目標(biāo)與其領(lǐng)域內(nèi)的背景關(guān)系缺乏考慮,尤其是當(dāng)目標(biāo)尺度大小發(fā)生重大變化或旋轉(zhuǎn)時(shí),容易跟丟目標(biāo)[20]。本文主要針對(duì)以上兩種情況,提出一種融合前景目標(biāo)空間信息的自適應(yīng)跟蹤算法,首先對(duì)初始幀感興趣目標(biāo)建立顏色直方圖作為目標(biāo)模型,然后在目標(biāo)框區(qū)域通過(guò)快速傅里葉變換得到頻譜圖,最后計(jì)算目標(biāo)模型與背景模型的置信度,得到目標(biāo)框調(diào)節(jié)因子,從而實(shí)現(xiàn)對(duì)目標(biāo)的自適應(yīng)跟蹤。