Face identification in near IR-videos using smart eye tracking profiles

Abstract

In this thesis we have evaluated the possibility of using data extracted from Smart Eye tracking pro les for the task of driver identi cation in near-IR videos using face recognition techniques. Two texture-based methods are presented to solve this task. The rst method utilizes local binary patterns (LBP) combined with local discriminant analysis (LDA) to identify subjects whilst the second method investigates brute-force template matching for identi cation purposes. The LBP-method was evaluated for both closed-set and open-set identi cation and achieved good results. In the closed-set scenario, a recognition rate of 98% was reached on our own data set of 47 videos containing 24 di erent subjects, despite the fact that the subjects had modi ed their appearances in half of the videos. The LBP-method is fast and our results demonstrates the LBP-method's ability to handle partial facial occulusions and appearance variations. The tem- plate matching method showed decent results with a peak recognition rate of 81% in the closed-set scenario. However, the method proved very time-consuming due to large number of templates in the pro les, limiting the method's viability in real-time applications. Future re nements and extensions are proposed for both methods.