Concept-Aware Ensemble System for Pedestrian Detection

Abstract

In ADAS, multiple classifier system on pedestrian detection is occupying important position because of its merit that each classifier can be able to create synergistic approaches to compensate the other member classifiers inability. On the other hand, according to different poses of pedestrians and variable background, once trained pedestrian detector needs to be tuned dynamically along the variation of real-world environment, thus the system is requested to incrementally accept new information and retain the old one at the same time. This thesis presents an incremental learning, environment-adaptive ensemble system for pedestrian detection by combining pedestrian detector constituted by multiple classifiers with front-end concept recognizer that can turn off inefficient member classifiers adaptively. Through adopting incremental learning algorithm, newly added batch dataset is trained by learning algorithm and the newly generated classifier is united to the existing ensemble along with the update of the voting weight. As the update of voting weight is only taken when the new training is carried out and focuses on the performance on current environment, temporal trade-off on performance between current and old environment is inevitable. This problem is addressed by applying concept recognizer in front of the ensemble thus turning off ineffective classifiers and selecting the most efficient voting weight vector adaptive to each pedestrian candidate. With the intervention of the front-end concept recognizer, the system can retain good performance on old environment while does not lose focus on current environment.