Human Action Recognition in Image Sequences

We (Esra Acar, Tobias Senst and Alexander Kuhn) presented our work on human action recognition titled as Human Action Recognition using Lagrangian Descriptors at MMSP 2012, Banff, Canada. In this work, we show that Lagrangian methods, besides their potential in crowd motion analysis, are well suited to model complex individual human activities in image sequences. We use two important Lagrangian descriptors to represent human actions:

Finite Time Lyapunov Exponents (FTLE) which describes motion boundaries and segments areas of different motion patterns.

Time-normalized Arc Length measure which corresponds to the accumulated average velocity at a certain point in the space time domain and represents the clustering of trajectories of similar flow geometry.

We fuse these Lagrangian descriptors at feature-level and use them in a linear SVM classification scheme. We evaluated our method on the Weizmann and KTH datasets. The action recognition accuracy results demonstrate that our approach is promising and that the performance is improved by fusing Lagrangian measures.

To learn more about our method, please refer to our paper Human Action Recognition using Lagrangian Descriptors.

(a) Reference image of frame 1, sequence wave2 (top) and walk (bottom) with the corresponding FTLE+ (b), FTLE- (c) and time-normalized arc length (d) field. (e-g) Average of the FTLE+, FTLE- and time-normalized arc length field for all corresponding sequences.

Performance Evaluation

We tested our method on two standard datasets: the Weizmann and the KTH datasets. The experiments showed that combining FTLE and time-normalized arc length measures improves the recognition accuracy and our results are comparable to state-of-the-art action recognition solutions. In the following tables, we present our recognition accuracy results.

Recognition accuracy results on the Weizmann and the KTH datasets

Sample Videos

We present here some video samples from the Weizmann and the KTH datasets showing the classification results of our method.

The Weizmann dataset:

The KTH dataset:

The Presentation at MMSP’2012

Leave a Reply

You can use these HTML tags

<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>