Pedestrian Detection with Spatially Pooled Features and Structured Ensemble Learning
Seminar Title: Pedestrian Detection with Spatially Pooled Features and Structured Ensemble Learning
Speaker: Dr. Sakrapee (Paul) Paisitkriangkrai , the Australia Taxation Office (ATO)
Date and Time: 11.00 AM – 12:00 pm, 14 August, 2015
Seminar Room: CB11.12.113
Seminar Chairman: Associate Professor Jian Zhang (jian.zhang@uts.edu.au)
Abstract: Complex Visual Tracking is an active research area in computer vision since it faces to some challenge problems which stimulate interesting of scientist, and also has many important applications, such as tracking of objects such as people, animal, vehicle, and the others, even including complex video events analysis. There are two main difficulties to overcome in this task: abrupt motion and non-rigidness of the tracked object. For the first problem, we have presented a novel solution, IASAMC, for the first time. IASAMC means Intensively Adaptive Stochastic Approach Mento Carlo sampling for tracking of abrupt motion, which can sample from the complicated filtering distribution. In sampling, the history samples of the simulated Markov chain are exploited to establish a density grid based predictive model, and the “optimal” transition kernel is adaptively learned by using this model to improve the mixing of MCMC sampling. More precisely, the proposed algorithm works based on two collaborative online learning processes, of which one aims to approximate to the target distribution by estimating the density of states (DoS), and another one attempts to estimate the optimal proposal distribution based on the density grid based predictive model. The proposed algorithm can effectively overcome the local-trap problem in sampling by random walk in the energy space, as well as speed up the convergence rate due to its adaptive sampling mechanism. Extensive experiments show that the proposed tracking algorithm works effectively and outperforms several state-of-the-art alternatives in tracking of various types of abrupt motions, including sudden dynamics changes, camera switching, low-frame-rate videos, and etc.
Short CV of Dr. Sakrapee (Paul) Paisitkriangkrai
Bio: Dr. Sakrapee (Paul) Paisitkriangkrai is a data mining scientist at the Australia Taxation Office (ATO) in Box Hill, Melbourne. Before joining the ATO, he was a postdoctoral researcher at The Australian Centre for Visual Technologies, The University of Adelaide. He received his Bachelor degree in computer engineering, the Master degree in biomedical engineering, and the PhD degree from the University of New South Wales, Sydney, Australia, in 2003 and 2010, respectively. His research interests include data mining, pattern recognition, image processing, and machine learning.