劉懿
關(guān)鍵詞: 目標(biāo)跟蹤; 遺傳算法; 運(yùn)動視頻; 粒子濾波; HSV分布模型; 退化權(quán)值
中圖分類號: TN713?34; TP391 ? ? ? ? ? ? ? ? ? ? 文獻(xiàn)標(biāo)識碼: A ? ? ? ? ? ? ? ? ? ?文章編號: 1004?373X(2019)03?0065?03
Abstract: As the mainstream technology of target tracking, particle filtering has broad application prospect in human motion video analysis. A motion video target tracking algorithm based on improved particle filtering model is proposed to further improve the accuracy of target tracking. The target observation model is constructed by using HSV distribution model, and then the particle filter and degradation weight are combined to detect whether the moving target appears in the target observation model. The genetic algorithm is introduced to improve the particle filtering algorithm, and eliminate the phenomenon of particle degradation. The test verification was conducted with the sports athlete video. The experimental results show that the proposed algorithm can effectively complete the human target tracking in motion video, and has higher accuracy and operation efficiency than other algorithms.
Keywords: target tracking; genetic algorithm; motion video; particle filtering; HSV distribution model; degeneration weight
目標(biāo)跟蹤技術(shù)最開始應(yīng)用于軍事領(lǐng)域,并逐漸在民用領(lǐng)域得到快速的推廣。目標(biāo)跟蹤技術(shù)能夠觀測被跟蹤目標(biāo)的屬性與狀態(tài),從而獲取被跟蹤目標(biāo)在不同時(shí)刻的變化。通過分析這些變化能夠?qū)δ繕?biāo)實(shí)現(xiàn)位置跟蹤[1?2]。一般來說,視頻目標(biāo)跟蹤需要對圖像序列進(jìn)行分析以便完成對運(yùn)動目標(biāo)的檢測,包括目標(biāo)的提取、識別和跟蹤,從而得到跟蹤目標(biāo)的各項(xiàng)運(yùn)動參數(shù),如加速度、速度、位置等[3]。
如何實(shí)現(xiàn)復(fù)雜背景下運(yùn)動目標(biāo)的準(zhǔn)確跟蹤一直是科研人員研究的熱點(diǎn)問題。基于蒙特卡羅思想的粒子濾波算法一直廣泛應(yīng)用于各種非線性及非高斯系統(tǒng),可以有效應(yīng)用于目標(biāo)跟蹤。因此,針對運(yùn)動視頻目標(biāo)跟蹤問題,本文提出一種基于改進(jìn)粒子濾波模型的運(yùn)動視頻目標(biāo)跟蹤算法。利用運(yùn)動員視頻進(jìn)行具體測試,結(jié)果顯示在無任何先驗(yàn)信息的情況下,提出的算法能夠較好地跟蹤運(yùn)行視頻中的人體目標(biāo),驗(yàn)證了其可行性和先進(jìn)性。
文獻(xiàn)[4]提出一種基于嵌入Mean?Shift的粒子濾波目標(biāo)跟蹤。文獻(xiàn)[5]提出面向顏色特征自適應(yīng)融合的改進(jìn)粒子濾波目標(biāo)跟蹤算法。文獻(xiàn)[6]提出基于粒子濾波和拉普拉斯方法的目標(biāo)跟蹤技術(shù)。以上幾種方法均采用混合優(yōu)化策略,通過將先進(jìn)的優(yōu)化算法和粒子濾波算法進(jìn)行結(jié)合來提高目標(biāo)跟蹤的性能,以便彌補(bǔ)粒子濾波算法的缺陷。遺傳算法作為一種仿生進(jìn)化式算法,其基本理念是適者生存規(guī)則和種群進(jìn)化,具有全局搜索能力高和前期收斂速度快的特點(diǎn),可用于消除粒子退化問題。因此,本文引入遺傳算法對粒子濾波算法進(jìn)行改進(jìn),以便增加粒子的多樣性,從而消除粒子退化的現(xiàn)象。此外,采用HSV分布模型構(gòu)建目標(biāo)觀測模型,然后結(jié)合粒子濾波器和退化權(quán)值檢測運(yùn)動目標(biāo)是否出現(xiàn)在目標(biāo)觀測模型中。
對三種跟蹤算法進(jìn)行測試,結(jié)果如表1所示。從表1可以看出,提出的方法明顯優(yōu)于其他兩種方法,其平均誤差精度一直維持在比較低的水平。在測試的視頻序列中,本文提出的跟蹤算法、標(biāo)準(zhǔn)粒子濾波算法、Mean?Shift粒子濾波算法的平均誤差分別為18.89,24.71,36.42。
本文提出一種基于改進(jìn)粒子濾波模型的運(yùn)動視頻目標(biāo)跟蹤算法。首先采用HSV分布模型構(gòu)建目標(biāo)觀測模型,然后結(jié)合粒子濾波器和退化權(quán)值來檢測運(yùn)動目標(biāo)是否出現(xiàn)在目標(biāo)觀測模型中。最后引入遺傳算法對粒子濾波算法進(jìn)行改進(jìn),以便消除粒子退化的現(xiàn)象。利用運(yùn)動員視頻進(jìn)行具體測試,結(jié)果顯示在無任何先驗(yàn)信息的情況下,提出算法能夠較好地跟蹤運(yùn)行視頻中的人體目標(biāo),驗(yàn)證了其可行性和先進(jìn)性。
參考文獻(xiàn)
[1] MILAN A, SCHINDLER K, ROTH S. Multi?target tracking by discrete?continuous energy minimization [J]. IEEE transactions on pattern analysis & machine intelligence, 2016, 38(10): 2054?2068.
[2] DEMIGHA O, HIDOUCI W K, AHMED T. On energy efficiency in collaborative target tracking in wireless sensor network: a review [J]. IEEE communications surveys & tutorials, 2013, 15(3): 1210?1222.
[3] YANG B, NEVATIA R. Multi?target tracking by online lear?ning a CRF model of appearance and motion patterns [J]. International journal of computer vision, 2014, 107(2): 203?217.
[4] 侯一民,賀子龍.嵌入Mean Shift的粒子濾波目標(biāo)跟蹤算法[J].計(jì)算機(jī)系統(tǒng)應(yīng)用,2012,21(12):80?84.
HOU Yimin, HE Zilong. Particle filter target tracking algorithm embedded in Mean Shift [J]. Computer systems and applications, 2012, 21(12): 80?84.
[5] BIAN L, LI T, WEI Y, et al. Improved particle filtering target tracking algorithm for HLBP and color feature adaptive fusion [J]. Journal of Nanjing Normal University, 2018(1): 45?49.
[6] QUANG P B, MUSSO C, GLAND F L. Particle filtering and the Laplace method for target tracking [J]. IEEE transactions on aerospace & electronic systems, 2016, 52(1): 350?366.
[7] ZHU S, WANG D, CHANG B L. Ground target tracking using UAV with input constraints [J]. Journal of intelligent & robotic systems theory & applications, 2013, 69(1): 417?429.
[8] HOANG H G, BA T V. Sensor management for multi?target tracking via multi?Bernoulli filtering [J]. Automatica, 2014, 50(4): 1135?1142.
[9] CHEN Y. Target tracking feature selection algorithm based on Adaboost [J]. Telkomnika Indonesian journal of electrical engineering, 2014, 12(1): 734?740.
[10] 呂韻秋,劉凱,費(fèi)聚鋒,等.基于壓縮跟蹤和遺傳算法的實(shí)時(shí)跟蹤方法[J].制導(dǎo)與引信,2016,37(4):34?39.
L? Yunqiu, LIU Kai, FEI Jufeng, et al. Real?time tracking method based on compression tracking and genetic algorithm [J]. Guidance & fuze, 2016, 37(4): 34?39.
[11] 劉峰,宣士斌,劉香品.基于云自適應(yīng)粒子群優(yōu)化粒子濾波的視頻目標(biāo)跟蹤[J].數(shù)據(jù)采集與處理,2015(2):452?463.
LIU Feng, XUAN Shibin, LIU Xiangpin. Video target tracking based on cloud adaptive particle swarm optimization particle filter [J]. Data acquisition and processing, 2015(2): 452?463.