Scalable Activity Recognition Framework using Ensemble Models
Annapurna P Patil, Anushri Srinath, Atif Adib, Dimple Doshi, Darshan Dalsaniya · 2019 Second International Conference on Advanced Computational and Communication Paradigms (ICACCP) · 2019
Recognizing Activities from raw video streams has been one of the most challenging problems in the domain of computer vision since past three decades. In this paper we propose a method of combining unsupervised and supervised methods to develop a scalable solution which can have real world applications in surveillance, video indexing, security and could help us have an efficient way of tackling large scale surveillance with minimal human intervention. With our unsupervised reconstruction CNN we can cluster different videos based on their dense frame representations and then train a separate classifier for each of the clusters. This methods allows us to easily add new activities to the system without training the entire system from scratch. For activity recognition among the clusters we use LSTM based Recurrent neural networks with different hyperparameter for each clusters' classifier.