- Title
- Feature selection for intelligent transportation systems
- Creator
- Peng, Yu
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2014
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- This thesis addresses the problem of feature selection for developing vision-based applications for intelligent transportation systems. The aim is to establish an efficient and robust framework for feature selection to guide the development of vision-based applications for intelligent transportation systems. Traditional traffic surveillance relies on monitoring by human operators. The developments of computer vision techniques and traffic cameras provide new solutions. Intelligent vision-based applications can work continuously without rest, which allows them to monitor and manage the traffic more efficiently than human beings. To understand the correlation and mechanism between these applications fully, we categorise all of the applications into four vision-based cognitive systems, specifically, vehicle, pedestrian, driver, and traffic infrastructure. Similar to the human cognitive system, cognitive systems in intelligent transportation systems recognise objects and events based on features. Feature selection is extremely critical for achieving good performance. However, this task presents a considerable challenge because an intelligent transportation system usually contains a complicated environment, multiple objects, and an unstable background. Previous studies have indicated that feature selection is usually performed at random and without convincing reasons. To address this problem, we originally propose an efficient framework for feature selection to guide the development of cognitive systems in intelligent transportation systems. More specifically, our framework includes the scheme of maximum dependency and minimum redundancy. The first scheme guides us to select a feature candidate set such as low-level, high-level, or hybrid features; next, the cognitive system targets such as vehicles, pedestrians, drivers, or roads with a system performance requirement that is evaluated in terms of the accuracy, processing speed and robustness. Subsequently, the scheme of minimum redundancy helps us to select the optimal feature subset from the candidate set by analysing the correlations among the group of features. In the remainder of this work, we focus on the integration of feature selection schemes into real-world cognitive systems for intelligent transportation. Vehicles are the primary road users. Vehicle surveillance is one of the most important components of an intelligent transportation system. Using the proposed framework of feature selection, a vehicle classification system with roadside cameras is proposed. As another main road user, pedestrians have different attributes, such as irregular shapes and frequent occlusions. A pedestrian detection and counting system is developed with the guidance of the framework of feature selection. Additionally, intelligent vehicles are important cognitive systems in an intelligent transportation system. With a monocular camera installed on the front of a vehicle, the front vehicle on the road is detected, and its pose is estimated simultaneously. This system progresses toward being a better driver assistant in the future. Finally, we propose an inference bag-of-features model that selects the optimal feature subset. We demonstrate the advances of this proposed model through testing on an extremely challenging task: gender classification based on face recognition. These aforementioned cognitive systems are the most important and essential components of intelligent transportation systems. Each system is a complex system rather than a single task. For example, in a vehicle type classification system, license plate detection and vehicle front extraction are considered together before classifying the types of vehicles. The proposed schemes for feature selection that are applied during the overall procedure for each system and the advantages of our feature selection schemes are shown next. Throughout this work, focussed emphasis is placed on performing a thorough performance evaluation for both the methodology and the real-world datasets. Several datasets that are used in this thesis have been made publicly available for further research in this field. Our results indicate that a significant improvement in performance is achieved by using our feature selection methods. This thesis concludes with a critical analysis of the work and an outlook for future research opportunities.
- Subject
- feature selection; intelligent transportation systems; machine learning; computer vision; image processing
- Identifier
- http://hdl.handle.net/1959.13/1043438
- Identifier
- uon:14187
- Rights
- Copyright 2014 Yu Peng
- Language
- eng
- Full Text
- Hits: 1005
- Visitors: 1421
- Downloads: 424
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | ATTACHMENT01 | Abstract | 555 KB | Adobe Acrobat PDF | View Details Download | ||
| View Details Download | ATTACHMENT02 | Thesis | 5 MB | Adobe Acrobat PDF | View Details Download |