From data to decision : boosting for time-constrained detection of objects in moving video
Date
2011
Authors
Overett, Gary Mark
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis considers the detection of humans and road signs in video with significant ego-motion. The aim is to work towards an automated vision system for real-time and low-latency detection of target objects from a vehicle moving at normal road speeds. Therefore, a strong emphasis is placed on both classifier robustness and the associated time-to-decision. A cascade of classifiers trained using the popular boosting algorithm, RealBoost, is adopted as the baseline system for our research, with novel improvements building upon this framework.
The approach attempts to follow the full throughput of such systems, starting from the input training data, including the construction of suitable learners, and ending with the final classifier decision criteria. Evaluation and testing is done on pedestrian and road sign datasets, which provide a testing comparison across a range of problem difficulties. The approach avoids assumptions about the ego-motion of the video and makes no class specific assumptions. Therefore, the resulting improvements are applicable to a wide variety of class detection problems and are useful in detection problems for both moving and static camera systems. Several contributions are made, offering improvements in terms of the error rate performance and evaluation speed of various classifiers.
- In Chapter 4 we consider the question of 'what makes a good dataset?' and develop a large pedestrian dataset for supervised learning (the NICTA Pedestrian Dataset).
- In Chapter 5 we propose two new weak learners, which improve per-feature discriminance and training characteristics.
- In Chapter 6 we develop a novel histogram of oriented gradients based feature type (LiteHOG PLUS), which has exceptionally favourable time-error object detection performance.
- In Chapter 7 we consider the issues relating to the pairing of various feature types with particular weak learners. The chapter provides a feature design guide to help researchers select an appropriate weak learner for a given feature.
- In Chapter 8 we consider three issues relating to the learning of classifiers with excellent time-error performance. Firstly, we demonstrate a tendency for boosting algorithms to select features in a biased fashion, favouring overfitting weak learners. Secondly, we test a recently proposed complexity-aware feature selection rule in conjunction with our new LiteHOG PLUS feature. Thirdly, we compare alternative boosting algorithms for performance.
- In Chapter 9 we consider the issues involved in the creation of boosted cascades and demonstrate the effects of varying cascade termination criteria on time-error performance.
Description
Keywords
Citation
Collections
Source
Type
Thesis (PhD)
Book Title
Entity type
Access Statement
Open Access
License Rights
Restricted until
Downloads
File
Description