Pedestrian detection by using weighted channel features with hierarchical region reduction

Abstract

This dissertation studies the problem of pedestrian detection in video applications. Identifying a pedestrian in real time to avoid collision for unmanned vehicles is mandatory. This problem can be mathematically transformed to the problem of classifying objects as pedestrian and non-pedestrian classes. The proposed framework is an appearance-based multichannel detectors with hierarchical pedestrian region reduction. It is intended for monocular environment since it is the simplest, least expensive, and most practical environment to work in real applications. The pedestrian is represented by the combination of channel features. The operation is broken down into two steps, namely, training and testing. In training step, the channel features are weighted by pedestrian template for meaningful feature selection process. Handling partial occlusion is carried out in testing step by constructing a hierarchical region reduction structure. Then a full pedestrian image is reduced to the horizontal and vertical regions. Each region is designed for specific level of occlusion, having its own region detector. All region detectors are combined in hierarchical fashion to reduce time and redundant processes. The experiment yields good results using standard benchmark dataset. The performance evaluation on miss rate, average precision, and the trade-off between running time and performance of the proposed detector show that the proposed detection framework is a reasonable option for real world applications.