Lagrangian-Convolutional Networks for Video Classification

在过去的几十年中，视频数据量迅速增加。这种发展提高了对有效方法的需求，以分析视频数据，而不是基于单帧图像的分析。具体地说，动态运动信息的检测、量化和分类是处理时间相关视频数据的一个关键方面，但仍需要充分利用。在这个项目中，我们要求资助开发一个基于拉格朗日方法的视频数据分析和分类的新系统。通过我们的初步工作，我们强调了使用这种方法结合最近的机器学习方法进行视频处理的创新潜力。通过这个项目，我们希望发展和充实这种方法，并旨在开发一种用于视频级描述和分类的新概念：拉格朗日卷积神经网络（LaCNN）。该概念利用了最近对视频序列中的运动签名的深入理解，并且更有效地利用了拉格朗日方法和编码的运动信息的能力。总之，与现有的最先进的方法相比，这种新颖的概念将导致具有竞争力的有效和更透明的视频分类系统。

The amount of video data has rapidly increased throughout the last decades. This development rises the demand for effective methods to analyse video data beyond a single frame image-based analysis. Specifically the detection quantification and classification of dynamic motion information is a crucial aspect for processing time-dependent video data which yet needs to be fully exploited. With this project we request funding to develop a novel system for analysis and classification of video data based on a Lagrangian methodology. With our preliminary work we highlighted the innovative potential for video processing using such methods in combination with recent machine learning approaches. With this project we like to evolve and substantiate this approach and aim to develop a novel concept for video-level description and classification: the Lagrangian- Convolutional Neural Network (LaCNN). This concept takes advantage of recent in-depth understanding of motion signatures in a video sequence and exploits the capabilities of the Lagrangian methodology and the encoded motion information more effectively. In summary this novel concept will lead to a competitive effective and more transparent video classification system in comparison to existing state of the art methods.