اﻟﮕﻮرﯾﺘﻢ ردﯾﺎﺑﯽ ﯾﺎدﮔﯿﺮي ﺗﺸﺨﯿﺺ ﺑﻬﺒﻮد داده ﺷﺪه ﺟﻬﺖ ﻧﺮخ ﻗﺎب ﭘﺎﯾﯿﻦ

An Improved Tracking-Learning-Detection Algorithm for Low Frame Rate

اﻟﮕﻮرﯾﺘﻢ ردﯾﺎﺑﯽ ﯾﺎدﮔﯿﺮي ﺗﺸﺨﯿﺺ )TLD( ﺳﻨﺘﯽ، ﻧﺴﺒﺖ ﺑﻪ ﭼﺎﻟﺶﻫﺎﯾﯽ ﻫﻤﭽﻮن ﺗﻐﯿﯿﺮات روﺷﻨﺎﯾﯽ، ﮐﻼﺗﺮﻫﺎ و ﻧﺮخ ﻗﺎب ﭘﺎﯾﯿﻦ ﺑﺴﯿﺎر ﺣﺴﺎس ﺑﻮده و ﺑﺎﻋﺚ ﺧﻄﺎ در ردﯾﺎﺑﯽ ﻫﺪف ﻣﯽﮔﺮدد. در راﺳﺘﺎي ﻏﻠﺒﻪ ﺑﺮ اﯾﻦ ﻣﺸﮑﻼت و ﺑﻬﺒﻮد ﻣﻘﺎوﻣﺖ اﻟﮕﻮرﯾﺘﻢ، ﻣﻌﻤﺎري اﻟﮕﻮرﯾﺘﻢ ردﯾﺎﺑﯽ ﯾﺎدﮔﯿﺮي ﺗﺸﺨﯿﺺ ﺑﺎ ﺗﺮﮐﯿﺐ اﻟﮕﻮرﯾﺘﻢ اﻧﺘﻘﺎل ﻣﺘﻮﺳﻂ و اﻟﮕﻮرﯾﺘﻢ ﯾﺎدﮔﯿﺮي ﻧﯿﻤﻪﻧﻈﺎرﺗﯽ ﻫﻢﯾﺎدﮔﯿﺮي، ﺑﻬﺒﻮد داده ﺷﺪه اﺳﺖ. اﯾﻦ ﺳﺎﺧﺘﺎر در ﺷﺮاﯾﻂ ﻧﺮخ ﻗﺎب ﭘﺎﯾﯿﻦ ﻧﺘﺎﯾﺞ ﺑﻬﺘﺮي را ﻧﺘﯿﺠﻪ ﻣﯽدﻫﺪ و ﻣﻘﺎوﻣﺖ و دﻗﺖ اﻟﮕﻮرﯾﺘﻢ را ﻧﺴﺒﺖ ﺑﻪ اﻟﮕﻮرﯾﺘﻢ ﺳﻨﺘﯽ ردﯾﺎﺑﯽ ﯾﺎدﮔﯿﺮي ﺗﺸﺨﯿﺺ اﻓﺰاﯾﺶ ﻣﯽدﻫﺪ. زﯾﺮا اﻟﮕﻮرﯾﺘﻢ ردﯾﺎﺑﯽ اﻧﺘﻘﺎل ﻣﺘﻮﺳﻂ ﻧﺴﺒﺖ ﺑﻪ ﭼﺮﺧﺶ، ﻣﻮاﻧﻊ ﺟﺰﺋﯽ، ﺗﻐﯿﯿﺮات اﻧﺪازه ﻣﻘﺎوم ﺑﻮده و ﺑﻪ ﺳﺎدﮔﯽ اﺟﺮا ﺷﺪه و ﺑﻪ ﻣﺤﺎﺳﺒﺎت ﮐﻤﯽ ﻧﯿﺎز دارد. از ﻃﺮف دﯾﮕﺮ اﻟﮕﻮرﯾﺘﻢ ﯾﺎدﮔﯿﺮي ﻧﯿﻤﻪﻧﻈﺎرﺗﯽ ﻫﻢﯾﺎدﮔﯿﺮي ﺑﺎ دو ﻃﺒﻘﻪﺑﻨﺪ ﻣﺴﺘﻘﻞ ﻣﯽﺗﻮاﻧﺪ ﺗﻐﯿﯿﺮات وﯾﮋﮔﯽﻫﺎي ﻫﺪف را ﺑﻪ ﺧﻮﺑﯽ آﻣﻮزش ﺑﺒﯿﻨﺪ. ﺑﻨﺎﺑﺮاﯾﻦ، ﺳﺎﺧﺘﺎر ﺗﻮﺳﻌﻪ داده ﺷﺪه ﻣﯽﺗﻮاﻧﺪ ﻣﺸﮑﻞ ﮔﻢ ﮐﺮدن ﻫﺪف را در ﺷﺮاﯾﻂ وﺟﻮد ﻫﻤﺰﻣﺎن ﻧﺮخ ﻗﺎب ﭘﺎﯾﯿﻦ و ﭼﺎﻟﺶﻫﺎي دﯾﮕﺮ ﺣﻞ ﻧﻤﺎﯾﺪ. ﻧﻬﺎﯾﺘﺎ، ارزﯾﺎﺑﯽ ﻣﻘﺎﯾﺴﻪاي روش ﭘﯿﺸﻨﻬﺎدي ﺑﺎ اﻟﮕﻮرﯾﺘﻢﻫﺎي ﻣﻌﺮوف ردﯾﺎﺑﯽ ﺑﺮ روي ﺳﻨﺎرﯾﻮﻫﺎي ﻣﺨﺘﻠﻒ از ﭘﺎﯾﮕﺎه داده ﻣﺸﻬﻮر 100-TB، ﺣﺎﮐﯽ از ﻋﻤﻠﮑﺮد ﺑﺮﺗﺮ روش ﭘﯿﺸﻨﻬﺎدي در ﻣﻘﺎﯾﺴﻪ ﺑﺎ ﺳﺎﯾﺮ روشﻫﺎ از ﻟﺤﺎظ ﻣﻘﺎوﻣﺖ و ﭘﺎﯾﺪاري اﺳﺖ. ﻧﻬﺎﯾﺘﺎ ﺳﺎﺧﺘﺎر ﭘﯿﺸﻨﻬﺎدي ﺑﺮ اﺳﺎس ﻣﻌﻤﺎري ردﯾﺎﺑﯽ ﯾﺎدﮔﯿﺮي ﺗﺸﺨﯿﺺ در وﯾﺪﯾﻮﻫﺎﯾﯽ ﺑﺎ ﭼﺎﻟﺶﻫﺎي ﻣﺨﺘﻠﻒ ذﮐﺮ ﺷﺪه ﺑﻪﻃﻮر ﻣﺘﻮﺳﻂ ﺣﺪود 33 درﺻﺪ ﻧﺘﺎﯾﺞ را ﻧﺴﺒﺖ ﺑﻪ اﻟﮕﻮرﯾﺘﻢ ﺳﻨﺘﯽ ردﯾﺎﺑﯽ ﯾﺎدﮔﯿﺮي ﺗﺸﺨﯿﺺ ﺑﻬﺒﻮد ﺧﻮاﻫﺪ ﺑﺨﺸﯿﺪ.

The conventional Tracking-Learning-Detection (TLD) algorithm is sensitive to illumination change and clutter and low frame rate and results in drift even missing. To overcome these shortcomings and increase robustness, by improving the TLD structure via integrating mean-shift and co-training learning can be achieved better results undergo low frame rate (LFR) condition and the robustness and accuracy tracking of the TLD structure increases. Because of, the Mean-Shift tracking algorithm is robust to rotation, partial occlusion and scale changing and it is simple to implement and takes less computational time. On the other, the co-training learning algorithm with two independent classifiers can learn changes of the target features in during the online tracking process. Therefore, the extended structure can solve the problem of lost object tracking in LFR videos and other challenges simultaneously. Finally, comparative evaluations of the proposed method to other top state-of-the-art tracking algorithms under the various scenarios from the TB-100 known dataset, demonstrate the superior performance of the proposed algorithm compared to other tracking algorithms in terms of tracking robustness and stability performance. Finally, the proposed structure based on the TLD architecture, in scenarios with the various challenges mentioned, will improve on average about 33% of the results, compared to the traditional TLD algorithm.