PARALLELYTICS · REAL-TIME VIDEO ANALYTICS ENGINE OPTIMIZED FOR GPUs

REAL-TIME VIDEO ANALYTICS ENGINE OPTIMIZED FOR GPUs

The objective of Parallelytics projects is to develop algoritms that can learn normalcy models of a scene from surveillance videos and detecting anormalities. Learning models from videos is a computationally complex task. Using longer video sequences for training models is desired to improve the learning strength. However, computational complexity increases rapidly, thereby enforcing suboptimal heuristic learning techniques and requiring significantly more hardware resources. Parallelytics is motivated by a few key observations. Due to the structure of video sequences, learning normalcy models from video can potentially be parallelized both temporally and spatially. This enables use of GPUs to achieve real-time anomaly detection. Furthermore, inference algorithms that are needed to estimate model parameters can be achieved by message-passing algorithms which maximize a probability distribution on the model parameters. Our outlined research goals in three main workpackages are as given below: • Parallelization friendly machine learning algorithm selection and design • Unsupervised and supervised learning algorithms and feature design for normalcy learning • Development of adaptive message-passing algorithms In the second half of the project work has been carried covering all three workpackages: • The machine learning course established by the researcher at the graduate level in Istanbul Sehir University has been incorporated into the undergraduate computer science curriculum. • We developed feature selection measures for unsupervised and supervised decision tree learning and dynamic time warping applications to modeling time series data. We applied these techniques to model social media discussion on platforms such as Twitter using real data. We also designed features for shape recognition, crowd activity recogniton, and invidual agent motion modeling. • Message-passing algorithms are used to maximize a probability distribution to estimate unknown variables. Oftentimes, these message-passing algorithms are regularized. We designed intelligent message-passing algorithms that does adaptive regularization and are parallelizable and GPU friendly. The second half of the project was fruitful for the researcher and the host organization. Two graduate level courses are created to establish the theoretical background for graduate students at Istanbul Sehir University: ECE 525 Machine Learning and ECE 528 Probabilistic Graphical Models. Furthermore, the researcher has established the Data Science Lab which has currently 11 graduate and undergraduate students. Data Science Lab has four other projects founded by Tubitak, the national research funding agency in Turkey. Two of these projects are proposed by the researcher. One of these projects is on social media analysis and has been selected to EU Cost Action on Computational Social Choice (IC1205). Furthermore, our first graduate student has been admitted to do his Ph. D studies in Computer Science department at Purdue University with full tuition-waiver and an assistantship. Our undergraduate reserach assistant has been admitted to do hos Ph.D studies in Computer Science department at Stony Brooks University with full tuition-waiver and scholarship. The project was very fruitful for the researcher and the host organization. The researcher was given the Associate Professor title by the Turkish Higher Education Councel. Alignment of the project with the general direction of the Computer Science department at Istanbul Sehir University and its emphasis on data science, has improved the impact of the project.

Data: CORDIS, © European Union