An Hybrid CPU-GPU Parallel Multi-tracking Framework for Long-Term Video Sequences

Abstract

The automatic evaluation of video content is today one of the biggest challenges in computer Vision. When the purpose is to work with static surveillance cameras, where most of the time the scenes do not change ,a full Convolutional Network (CNN) approach seems to require too much CPU effort, specially when the objects are slightly moving between different frames. On the other side, visual tracking has seen great recent advances in either speed or accuracy but still remain scarce when have to deal with long videos where objects constantly new ones come into the scene and others disappear. In this paper, we present a parallelization scheme to handle multiple instances of object tracking. The main purpose is reduce overall processing time . The idea is to use already pre-trained CNNs for discovering objects and a parallel multi-tracker for following them, using both CPU and GPU devices. Our multi-tracker framework consists of three main components, a movement detector, an object classification and a tracker. We use the object detector as an initialization for trackers. When there are plenty of objects in the scene, the other two components are incorporated for reducing CPU effort. The first one is a scheduler than prioritizes tracking those objects that seems more relevant than the others. This scheduler use a criteria that balances the multi-tasking trying to reach the greatest speed-up with minimal detections lost. The second one, is a GPU memory handler, that lets adapt the framework to different hardware configuration specially when the CNNs could not be completely allocated into the device. As a general framework, it is very flexible and it could be customized with different trackers and CNN, adapting to different situations and platforms. We evaluate this framework in different cases and cameras configurations, reaching reasonable speed-up and confidence.