Weighted template construction for pedestrian detection using biased boosting

Detecting pedestrian is an important step in many areas, such as intelligent transportation systems (ITSs) or visual surveillance. Currently, boosting a set of local features based on histogram of oriented gradients (HOGs) to form a pedestrian detector has proven its effectiveness in the literature. However, this kind of approaches suffer from the problem of false detection in case of complex background or noise effect. Accordingly, the main objective of this work is to alleviate this problem by integrating the results from the template matching, which is a kind of global features into the boosting framework. The idea behind is to adjust the hyperplane of the support vector machine according to the template-based classifier at each round of boosting stage. This makes both global and local features complement each other and the learned detector raises the detection rate and reduces the false positive rate at the same time. Instead of manual annotation, a set of representative templates are automatically constructed based on expectation maximization (EM) algorithm. To make the template have more discriminative power, we assign each point in the constructed template a different weight in matching but not consider all points as equally important. The experiments provided exhibit the superiority of the proposed method in detection accuracy.