Quadratic binary optimization for pedestrian detection in crowded scenes
Pedestrian detection is of much interest in the computer vision research community and is a rapidly evolving research area. Significant progress has been demonstrated but the performance degrades rapidly when occlusion occurs. We propose a binary quadratic optimization framework for multi-pedestrian...
Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Dissertation |
| Sprache: | Englisch |
| Veröffentlicht: |
ProQuest Dissertations & Theses
01.01.2014
|
| Schlagworte: | |
| ISBN: | 9781369405132, 1369405138 |
| Online-Zugang: | Volltext |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | Pedestrian detection is of much interest in the computer vision research community and is a rapidly evolving research area. Significant progress has been demonstrated but the performance degrades rapidly when occlusion occurs. We propose a binary quadratic optimization framework for multi-pedestrian detections to improve pedestrian detection, especially when there is an occlusion. The pedestrian detection problem is formulated as a Quadratic Unconstrained Binary Optimization (QUBO) problem. The formulation reasons directly over the space of overlapping among object detections as a pairwise measurement and an individual detection confidence score as a unary measurement. The tradeoff between detection confidences and amounts of overlap is optimized thereby allowing a more relaxed selection compared to conventional non-maximum suppression approach. Although QUBO is an NP-hard problem, efficient approximate methods are available, and these yield high quality solutions on large problem sizes. The core concept of optimized detection framework is to generate a large set of possible detection candidates and then use the optimization method to select a subset of those candidates that best represents the detection. The optimized detection framework is further applied to a multiple-body-part representation called "body plan". With the adoption of multiple parts, reasoning about occlusions among pairs of overlapped detections, by applying a positive (reward) and negative (penalty) pairwise scores, can help to decide whether the missing body part is likely due to being occluded. Besides detection candidates, it is also applicable to use tracklets of matching as candidates and formulate the problem as non bipartite matching graph. We reformulate the non bipartite matching problem into a Quadratic Constrained Binary Optimization that solves for a set of detections along with data association. The proposed optimized detection framework shows that quadratic optimization for reasoning about overlapping detections and quality of an individual detection can improve the performance of a pedestrian detection system, especially when there are multiple occluded persons. We also show that the framework of optimized detection can be applied to several types of candidates, not only a sliding window based detector but also shape covering and pairs of matching candidates. |
|---|---|
| Bibliographie: | SourceType-Dissertations & Theses-1 ObjectType-Dissertation/Thesis-1 content type line 12 |
| ISBN: | 9781369405132 1369405138 |