Analyzing low light adaption methods for pre-trained cascade of boosted MB-LBP classifiers

Sammanfattning: In applications of real time gesture detection on devices which have cameras and limited power supply it is important to consider both robustness and maintaining a low power consumption. A well explored approach is the use of cascades of boosted classifiers. As retraining is a very cumbersome process, online environment adaption would be an appealing approach to counter situations where the boosted cascade detector underperforms. The purpose of this thesis is to investigate ways to adapt a pre-trained boosted cascade detector, with regression trees as weak classifiers, to an elusive environment in test-time. The approach is unsupervised collection of data in real time on highly probable false negatives coming from temporal series of frames. The data is then used to find features associated with missed detections and increase their influence on the classifier while keeping the unwanted impact bounded. Adaption is accomplished by reclassifying any new low confidence data with a set of adapted leaves. In this thesis, adaption to a low light environment is investigated. Methods to unsupervised collect data and output adapted leaf values are presented. The results suggest convergence of the increased detection rates up to 13%, with respect to both the number of target stages and the size of the reclassification area, can be achieved with a few percent increase of feature evaluations. In order to draw further conclusions, about the general performance as well as the environment biased performance, more data is needed together with appropriate tools for analyzing results, such as ROC-curves.