對光線變化具有強健適應的人物重新識別系統輔以基於群聚的損失函數

Abstract

近年來，人物重新識別系統受到大量的關注，因為其有廣大的應用場域像是智慧家庭、健康照護以及監視系統。但是隨著視角的改變以及拍攝相機的位置不同，人的輪廓外觀也會跟著不同，這造成了從不同的視角進行行人追蹤仍然是個挑戰。 除此之外，在實際應用的場域中在各個相機之間的光線照射位置及程度都是不同的，而在當前的人物重新識別模組中往往只會透過現有的資料集來學習這使得模型不具有能應付光亮變化的能力。最後我們在分析了當前最先進的人物重新識別期刊文章後我們也發現，目前大多數的研究為了要訂出閾值來區分正樣本及負樣本，都會採用指標損失函數來做模型優化，雖然指標損失函數可以有效的區分正負樣本，但使用這種損失函數卻需要面臨因時間複雜度過高而使得訓練冗長，導致我們將模型移轉到新環境時即使已經蒐集好新環境的資料仍然需要花費大量時間進行模型校正。 首先若要解決光亮變化的問題，最直觀的方式就是搜集大量光亮變化的人物重新識別資料集，但是這件事情卻是相當困難且需耗費相當多時間以及人力的。基於上述原因，本論文提出了一種透過合成資料來協助模型訓練以提取無關光亮變化的特徵向量。而針對指標損失函數時間複雜度過高的缺點，本研究也提出了一種基於群聚的損失函數以降低時間複雜度並且效能更優於指標損失函數。。並在最終實驗也證明本論文提出的方法及損失函數在人物重新識別的任務中超過其他行人重新識別方法。

Nowadays, person re-identification has raised lots of attention in the area of computer vision, because of its wide applications, including smart home, elderly care, and surveillance systems. From different viewpoints, the shape of the human body looks completely different, hence tracking human from different camera remains a challenging problem. In addition, the locations and levels of light illumination can be different among cameras in the field of actual application. However, the existing person re-identification module is often learned only through the available dataset, which make model fail to be robust to situations with illumination change. Finally, after analyzing the recent literature on pedestrian re-recognition, we also found that most of the current researches use the metric loss function to optimize the model with an appropriate threshold to distinguish the positive samples from the negative ones. Despite the metric loss function can perform objective distinction as mentioned, it remains to have a disadvantage of having highly complex, which make the training process lengthy. First of all, to solve the problem of brightness changes, the most intuitive way is to collect an even larger person re-identification dataset subject to various brightness levels, which however is very expensive to collect and label. Therefore, this thesis proposes an illumination-invariant feature vector that assists model training based on synthetic data. To remove the shortcomings of the time complexity of the metric loss function, we propose the clustering-based loss function to reduce the time complexity, and we also show that the performance of the proposed loss function is better than metric loss function. In the final experiment, it is also proved that the proposed method in this thesis excels the state-of-the-art methods on resolving person re-identification problems.