Improved Boosting Algorithms Using Confidence-rated Predictions
Machine Learning - The Eleventh Annual Conference on computational Learning Theory
Histograms of Oriented Gradients for Human Detection
CVPR '05 Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) - Volume 1 - Volume 01
Multi-Object Tracking Through Simultaneous Long Occlusions and Split-Merge Conditions
CVPR '06 Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 1
Robust People Tracking with Global Trajectory Optimization
CVPR '06 Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 1
Fast Human Detection Using a Cascade of Histograms of Oriented Gradients
CVPR '06 Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2
International Journal of Computer Vision
Localizing Objects with Smart Dictionaries
ECCV '08 Proceedings of the 10th European Conference on Computer Vision: Part I
Robust Object Tracking by Hierarchical Association of Detection Responses
ECCV '08 Proceedings of the 10th European Conference on Computer Vision: Part II
| Hi-index | 0.00 |
We present a framework addressing the problem of multi-pedestrian tracking in a crowded scene from a single view camera. The key factor to improve the performance of tracking results in data association is providing an effective appearance model for each target. There are many efforts in developing such models in generative and discriminative ways to either describe the object well or discriminate it from others. Here we propose an efficient algorithm to learn both discriminative and generative appearance models online for different targets in a hierarchical framework. Thus, we are able to emphasize the similarity between a current tracklet (short track segment) and its new association target while differentiate itself from others. Within a time sliding window, detection responses are gathered to form short track fragments based on spatial-temporal information. These tracklets provide us affinity and discriminative information, which can be used to collect samples for learning models. To learn appearance model, we use AdaBoost to maximize discriminative capability. Kalman filter is also used to model the motion as an important factor to associate tracklets together. In the experiment, we evaluate different features setup of our framework on CAVIAR public dataset. It shows that our method can successfully track in heavy occluded scenes. It also helps improve tracking accuracy by overcoming missed detections, false alarms and ID switching issue. We compare our method with several state-of-the-art algorithms for evaluation.