 |
Publications |
|
|
|
|
|
|
 Katz, R. |
|
Track Based Classification of Dynamic Obstacles
PhD thesis, The University of Sydney, Aug, 2009
|
Abstract
This thesis is concerned with the unsupervised classification of dynamic obstacles in urban environments using 2D laser and visual data. Reliable classification of the surrounding obstacles is crucial for autonomous vehicles navigating in cities and assistance systems for drivers. Classification of dynamic obstacles in urban scenarios is a particularly challenging task. Perception must deal with different types of objects with varying dynamics and levels of occlusion, in scenarios affected by changing lighting and environmental conditions.
In order to address these challenges a track-based model is proposed that integrates laser and visual information for spatio-temporal synthesis of the sensed moving obstacles. This forms the basis for algorithms to perform unsupervised classification by clustering. Various contributions are made in order to achieve accurate and efficient performance, initially using laser and then incorporating vision into the process.
Firstly, an approach for the representation of laser tracks referred to as laser stamps is introduced, together with a measure of similarity. This allows the system to robustly capture shape information from the sensed dynamic obstacles and to be able to distinguish them. An extension of the affinity propagation (AP) algorithm for incremental clustering is proposed, which permits efficient on-line grouping of large numbers of laser stamps through the associated similarities. The algorithms for unsupervised classification based on laser are validated using a simulated environment that produces laser data, achieving accuracies of over 95% in three-class and nine-class scenarios.
Secondly, visual information is integrated to the system to improve the classification by effectively reducing the number of clusters. A formulation is introduced to represent visual tracks through visual stamps using a single-instance feature-based approach. A visual similarity learning method based on positive-only learning (PL) is proposed to compute similarity between the visual stamps based on a priori clustering given by the laser. Laser and vision combination is then achieved by introducing a combined similarity measure.
For accurate combined clustering an algorithm that integrates this combined similarity within an iterative clustering algorithm based on AP is derived. The combined laser and vision approach is validated for unsupervised classification of dynamic obstacles, using the simulated environment producing laser data and integrating images from a public labeled database (LabelMe). Experiments show that the system is able to maintain an accuracy of over 94%, while notably reducing the number of clusters.
Lastly, the architecture is evaluated using data collected by an experimental vehicle equipped with a 2D laser and a color monocular camera navigating in an urban environment. To deal with challenges associated with perception in real-world scenarios, an extension of the visual similarity learning approach based on multiple instance learning (MIL) is proposed. This extended model considers visual stamps derived from the full sequences of images rather than a single-instance representation for the tracks, and permits the direct integration of the framework to robustly deal with track-based laser and visual information. The final experimental evaluation in real-world urban environments demonstrates the performance of the proposed methodology, which reaches an accuracy of over 92% and finds the 3 clusters that correspond to the main obstacle classes in the data.
|
| Download |
PDF full text (23.50M)
|
|
Copyright Notice & Disclaimer |
|