SlideShare a Scribd company logo
Human Action Recognition using Lagrangian Descriptors

Esra Acar, Tobias Senst, Alexander Kuhn, Ivo Keller, Holger Theisel, Sahin Albayrak, Thomas Sikora



Esra Acar
Competence Center Information Retrieval and Machine Learning

                            VideoSense Project (www.videosense.eu)
                               SemSeg Project (www.semseg.org )
Outline

▶ Introduction
▶ Lagrangian Methods
▶ Our Method
   Lagrangian Descriptors
   Lagrangian Classifier
   Performance Evaluation
▶ Privacy Aspect
▶ Conclusions & Future Work




 MMSP’2012      Human Action Recognition using Lagrangian Descriptors   2
Introduction

▶ Human action recognition
   pertains to a multitude of application domains, ranging
      from automatic video surveillance to semantic video
      indexing and retrieval.
   requires the description of complex motion patterns in
      image sequences.
▶ Our aim in this work: To provide a reliable solution for
  recognizing basic human actions such as walking, running
  and boxing.
▶ Final goal: To build a solution to recognize more complex
  actions, in particular in the scenarios involving more than
  one basic action.


 MMSP’2012       Human Action Recognition using Lagrangian Descriptors   3
Lagrangian Methods

▶ When the sequence of an optical flow fields is treated as an
  unsteady vector field, then Lagrangian analysis can be
  applied.
▶ Lagrangian methods have been used for crowd
  segmentation.

▶ The concept of Lagrangian Coherent Structures (LCS)
  describe the properties of trajectories in the space time
  domain.
▶ In video analysis, this trajectory description corresponds to
  the notion of an edge of an enclosed moving object within
  the image.


 MMSP’2012        Human Action Recognition using Lagrangian Descriptors   4
Our Method

▶ We focus on two important Lagrangian motion features for
  the description of human actions in image sequences:

       Motion boundaries denote areas of high particle trajectory
        separation and segment areas of different motion patterns
        of varying time intervals.

       Areas of coherent flow motion gather regions of similar
        speed over the respective time interval.




 MMSP’2012         Human Action Recognition using Lagrangian Descriptors   5
Lagrangian Descriptors - 1




▶ The flow map describes a mapping of an initial position to its
  next position after a predefined integration time τ starting at
  t0.
▶ Combining all positions of one specific point within a certain
  time interval [t0; t0 + τ] creates a polynomial curve denoted
  as path line.

 MMSP’2012        Human Action Recognition using Lagrangian Descriptors   6
Lagrangian Descriptors - 2

▶ Technique to extract LCS: Finite-Time Lyapunov Exponents
  (FTLE) which is a method that
   describes the motion boundaries, and
   separates regions of coherent flow behavior.
▶ Given the optical flow field, FTLE is computed as follows:




 MMSP’2012       Human Action Recognition using Lagrangian Descriptors   7
Lagrangian Descriptors - 3

▶ Similar motion behavior can be formulated using further
  geometric features of the path lines.

▶ The areas of similar motion behaviors over time are
  represented with time-normalized arc length (Ʌ ).
                                                arcL


▶ The time-normalized arc length of a given time t0 and an
  integration interval τ is defined as:




 MMSP’2012      Human Action Recognition using Lagrangian Descriptors   8
Lagrangian Descriptors - 4




 MMSP’2012     Human Action Recognition using Lagrangian Descriptors   9
Lagrangian Descriptors - 5

▶ We compute Histogram of oriented Gradients (HoG) to avoid
  explicit extraction of ridge structures.




 MMSP’2012      Human Action Recognition using Lagrangian Descriptors   10
Lagrangian Classifier

▶ FTLE-HoG and Ʌ -HoG descriptors of video frames provide
                 arcL
  complementary information about an action.
▶ These two descriptors are fused before training linear multi-
  class SVMs.




 MMSP’2012        Human Action Recognition using Lagrangian Descriptors   11
Performance Evaluation - 1

             Results on the Weizmann dataset




                Results on the KTH dataset




 MMSP’2012       Human Action Recognition using Lagrangian Descriptors   12
Performance Evaluation - 2
   Confusion Matrix on the Weizmann dataset

                                               Most confused action pair is
                                               skip-jump pair.




   Confusion Matrix on the KTH dataset

                                                Most confused action pairs are
                                                • jogging-running,
                                                • jogging-walking,
                                                • handclapping-hand waving,
                                                • boxing-hand clapping.



 MMSP’2012       Human Action Recognition using Lagrangian Descriptors           13
Privacy Aspect

▶ Working with image data introduces another important
  aspect which is the privacy of the recorded individuals.

▶ In our method, we do not store image sequences directly,
  but optical flow fields.

▶ Optical flow fields only encode abstract motion patterns.

▶ The recognition of person-related features extracted from
  abstract Lagrangian motion patterns is much more difficult.




 MMSP’2012        Human Action Recognition using Lagrangian Descriptors   14
Conclusions & Future Work

▶     This study was a starting point to represent & recognize complex
      human actions.
▶     The main difference from previous works is that our method
      transforms the motion information of an individual in a given time
      interval into a 2D space.
▶     It is shown that FTLE and Ʌ Lagrangian measures are well
                                     arcL
      adapted to model individual human activities.

▶     Future Work
       Local motion representation – Interest points-based method.
       Multi-level motion representation with multi-level Bag-of-Words
         approach.
       Recognizing more complex human motions.


    MMSP’2012         Human Action Recognition using Lagrangian Descriptors   15
Thanks!

                  Questions?




MMSP’2012   Human Action Recognition using Lagrangian Descriptors   16

More Related Content

Similar to Human Action Recognition using Lagrangian Descriptors

HUMAN ACTION RECOGNITION IN VIDEOS USING STABLE FEATURES
HUMAN ACTION RECOGNITION IN VIDEOS USING STABLE FEATURES HUMAN ACTION RECOGNITION IN VIDEOS USING STABLE FEATURES
HUMAN ACTION RECOGNITION IN VIDEOS USING STABLE FEATURES
sipij
 
Human Action Recognition Using Deep Learning
Human Action Recognition Using Deep LearningHuman Action Recognition Using Deep Learning
Human Action Recognition Using Deep Learning
IRJET Journal
 
Sparse representation based human action recognition using an action region-a...
Sparse representation based human action recognition using an action region-a...Sparse representation based human action recognition using an action region-a...
Sparse representation based human action recognition using an action region-a...
Wesley De Neve
 
Gait Recognition using MDA, LDA, BPNN and SVM
Gait Recognition using MDA, LDA, BPNN and SVMGait Recognition using MDA, LDA, BPNN and SVM
Gait Recognition using MDA, LDA, BPNN and SVM
IJEEE
 
Human Action Recognition in Videos Employing 2DPCA on 2DHOOF and Radon Transform
Human Action Recognition in Videos Employing 2DPCA on 2DHOOF and Radon TransformHuman Action Recognition in Videos Employing 2DPCA on 2DHOOF and Radon Transform
Human Action Recognition in Videos Employing 2DPCA on 2DHOOF and Radon Transform
Fadwa Fouad
 
Silhouette analysis based action recognition via exploiting human poses
Silhouette analysis based action recognition via exploiting human posesSilhouette analysis based action recognition via exploiting human poses
Silhouette analysis based action recognition via exploiting human poses
AVVENIRE TECHNOLOGIES
 
Human Action Recognition
Human Action RecognitionHuman Action Recognition
Human Action Recognition
NAVER Engineering
 
Survey on Human Behavior Recognition using CNN
Survey on Human Behavior Recognition using CNNSurvey on Human Behavior Recognition using CNN
Survey on Human Behavior Recognition using CNN
IRJET Journal
 
Temporal Reasoning Graph for Activity Recognition
Temporal Reasoning Graph for Activity RecognitionTemporal Reasoning Graph for Activity Recognition
Temporal Reasoning Graph for Activity Recognition
IRJET Journal
 
Activity Recognition Using RGB-Depth Sensors-Final report
Activity Recognition Using RGB-Depth Sensors-Final reportActivity Recognition Using RGB-Depth Sensors-Final report
Activity Recognition Using RGB-Depth Sensors-Final report
nazlitemu
 
Crowd Recognition System Based on Optical Flow Along with SVM classifier
Crowd Recognition System Based on Optical Flow Along with SVM classifierCrowd Recognition System Based on Optical Flow Along with SVM classifier
Crowd Recognition System Based on Optical Flow Along with SVM classifier
IJECEIAES
 
IRJET- Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolu...
IRJET- Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolu...IRJET- Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolu...
IRJET- Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolu...
IRJET Journal
 
A Survey On Tracking Moving Objects Using Various Algorithms
A Survey On Tracking Moving Objects Using Various AlgorithmsA Survey On Tracking Moving Objects Using Various Algorithms
A Survey On Tracking Moving Objects Using Various Algorithms
IJMTST Journal
 
IRJET- Recognition of Human Action Interaction using Motion History Image
IRJET-  	  Recognition of Human Action Interaction using Motion History ImageIRJET-  	  Recognition of Human Action Interaction using Motion History Image
IRJET- Recognition of Human Action Interaction using Motion History Image
IRJET Journal
 
Wearable sensor-based human activity recognition with ensemble learning: a co...
Wearable sensor-based human activity recognition with ensemble learning: a co...Wearable sensor-based human activity recognition with ensemble learning: a co...
Wearable sensor-based human activity recognition with ensemble learning: a co...
IJECEIAES
 
Dd net
Dd netDd net
Dd net
Fan Yang
 
Pedestrian behavior/intention modeling for autonomous driving II
Pedestrian behavior/intention modeling for autonomous driving IIPedestrian behavior/intention modeling for autonomous driving II
Pedestrian behavior/intention modeling for autonomous driving II
Yu Huang
 
CHARACTERIZING HUMAN BEHAVIOURS USING STATISTICAL MOTION DESCRIPTOR
CHARACTERIZING HUMAN BEHAVIOURS USING STATISTICAL MOTION DESCRIPTORCHARACTERIZING HUMAN BEHAVIOURS USING STATISTICAL MOTION DESCRIPTOR
CHARACTERIZING HUMAN BEHAVIOURS USING STATISTICAL MOTION DESCRIPTOR
sipij
 
IRJET- Behavior Analysis from Videos using Motion based Feature Extraction
IRJET-  	  Behavior Analysis from Videos using Motion based Feature ExtractionIRJET-  	  Behavior Analysis from Videos using Motion based Feature Extraction
IRJET- Behavior Analysis from Videos using Motion based Feature Extraction
IRJET Journal
 
Pedestrian behavior/intention modeling for autonomous driving III
Pedestrian behavior/intention modeling for autonomous driving IIIPedestrian behavior/intention modeling for autonomous driving III
Pedestrian behavior/intention modeling for autonomous driving III
Yu Huang
 

Similar to Human Action Recognition using Lagrangian Descriptors (20)

HUMAN ACTION RECOGNITION IN VIDEOS USING STABLE FEATURES
HUMAN ACTION RECOGNITION IN VIDEOS USING STABLE FEATURES HUMAN ACTION RECOGNITION IN VIDEOS USING STABLE FEATURES
HUMAN ACTION RECOGNITION IN VIDEOS USING STABLE FEATURES
 
Human Action Recognition Using Deep Learning
Human Action Recognition Using Deep LearningHuman Action Recognition Using Deep Learning
Human Action Recognition Using Deep Learning
 
Sparse representation based human action recognition using an action region-a...
Sparse representation based human action recognition using an action region-a...Sparse representation based human action recognition using an action region-a...
Sparse representation based human action recognition using an action region-a...
 
Gait Recognition using MDA, LDA, BPNN and SVM
Gait Recognition using MDA, LDA, BPNN and SVMGait Recognition using MDA, LDA, BPNN and SVM
Gait Recognition using MDA, LDA, BPNN and SVM
 
Human Action Recognition in Videos Employing 2DPCA on 2DHOOF and Radon Transform
Human Action Recognition in Videos Employing 2DPCA on 2DHOOF and Radon TransformHuman Action Recognition in Videos Employing 2DPCA on 2DHOOF and Radon Transform
Human Action Recognition in Videos Employing 2DPCA on 2DHOOF and Radon Transform
 
Silhouette analysis based action recognition via exploiting human poses
Silhouette analysis based action recognition via exploiting human posesSilhouette analysis based action recognition via exploiting human poses
Silhouette analysis based action recognition via exploiting human poses
 
Human Action Recognition
Human Action RecognitionHuman Action Recognition
Human Action Recognition
 
Survey on Human Behavior Recognition using CNN
Survey on Human Behavior Recognition using CNNSurvey on Human Behavior Recognition using CNN
Survey on Human Behavior Recognition using CNN
 
Temporal Reasoning Graph for Activity Recognition
Temporal Reasoning Graph for Activity RecognitionTemporal Reasoning Graph for Activity Recognition
Temporal Reasoning Graph for Activity Recognition
 
Activity Recognition Using RGB-Depth Sensors-Final report
Activity Recognition Using RGB-Depth Sensors-Final reportActivity Recognition Using RGB-Depth Sensors-Final report
Activity Recognition Using RGB-Depth Sensors-Final report
 
Crowd Recognition System Based on Optical Flow Along with SVM classifier
Crowd Recognition System Based on Optical Flow Along with SVM classifierCrowd Recognition System Based on Optical Flow Along with SVM classifier
Crowd Recognition System Based on Optical Flow Along with SVM classifier
 
IRJET- Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolu...
IRJET- Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolu...IRJET- Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolu...
IRJET- Human Fall Detection using Co-Saliency-Enhanced Deep Recurrent Convolu...
 
A Survey On Tracking Moving Objects Using Various Algorithms
A Survey On Tracking Moving Objects Using Various AlgorithmsA Survey On Tracking Moving Objects Using Various Algorithms
A Survey On Tracking Moving Objects Using Various Algorithms
 
IRJET- Recognition of Human Action Interaction using Motion History Image
IRJET-  	  Recognition of Human Action Interaction using Motion History ImageIRJET-  	  Recognition of Human Action Interaction using Motion History Image
IRJET- Recognition of Human Action Interaction using Motion History Image
 
Wearable sensor-based human activity recognition with ensemble learning: a co...
Wearable sensor-based human activity recognition with ensemble learning: a co...Wearable sensor-based human activity recognition with ensemble learning: a co...
Wearable sensor-based human activity recognition with ensemble learning: a co...
 
Dd net
Dd netDd net
Dd net
 
Pedestrian behavior/intention modeling for autonomous driving II
Pedestrian behavior/intention modeling for autonomous driving IIPedestrian behavior/intention modeling for autonomous driving II
Pedestrian behavior/intention modeling for autonomous driving II
 
CHARACTERIZING HUMAN BEHAVIOURS USING STATISTICAL MOTION DESCRIPTOR
CHARACTERIZING HUMAN BEHAVIOURS USING STATISTICAL MOTION DESCRIPTORCHARACTERIZING HUMAN BEHAVIOURS USING STATISTICAL MOTION DESCRIPTOR
CHARACTERIZING HUMAN BEHAVIOURS USING STATISTICAL MOTION DESCRIPTOR
 
IRJET- Behavior Analysis from Videos using Motion based Feature Extraction
IRJET-  	  Behavior Analysis from Videos using Motion based Feature ExtractionIRJET-  	  Behavior Analysis from Videos using Motion based Feature Extraction
IRJET- Behavior Analysis from Videos using Motion based Feature Extraction
 
Pedestrian behavior/intention modeling for autonomous driving III
Pedestrian behavior/intention modeling for autonomous driving IIIPedestrian behavior/intention modeling for autonomous driving III
Pedestrian behavior/intention modeling for autonomous driving III
 

Human Action Recognition using Lagrangian Descriptors

  • 1. Human Action Recognition using Lagrangian Descriptors Esra Acar, Tobias Senst, Alexander Kuhn, Ivo Keller, Holger Theisel, Sahin Albayrak, Thomas Sikora Esra Acar Competence Center Information Retrieval and Machine Learning VideoSense Project (www.videosense.eu) SemSeg Project (www.semseg.org )
  • 2. Outline ▶ Introduction ▶ Lagrangian Methods ▶ Our Method  Lagrangian Descriptors  Lagrangian Classifier  Performance Evaluation ▶ Privacy Aspect ▶ Conclusions & Future Work MMSP’2012 Human Action Recognition using Lagrangian Descriptors 2
  • 3. Introduction ▶ Human action recognition  pertains to a multitude of application domains, ranging from automatic video surveillance to semantic video indexing and retrieval.  requires the description of complex motion patterns in image sequences. ▶ Our aim in this work: To provide a reliable solution for recognizing basic human actions such as walking, running and boxing. ▶ Final goal: To build a solution to recognize more complex actions, in particular in the scenarios involving more than one basic action. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 3
  • 4. Lagrangian Methods ▶ When the sequence of an optical flow fields is treated as an unsteady vector field, then Lagrangian analysis can be applied. ▶ Lagrangian methods have been used for crowd segmentation. ▶ The concept of Lagrangian Coherent Structures (LCS) describe the properties of trajectories in the space time domain. ▶ In video analysis, this trajectory description corresponds to the notion of an edge of an enclosed moving object within the image. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 4
  • 5. Our Method ▶ We focus on two important Lagrangian motion features for the description of human actions in image sequences:  Motion boundaries denote areas of high particle trajectory separation and segment areas of different motion patterns of varying time intervals.  Areas of coherent flow motion gather regions of similar speed over the respective time interval. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 5
  • 6. Lagrangian Descriptors - 1 ▶ The flow map describes a mapping of an initial position to its next position after a predefined integration time τ starting at t0. ▶ Combining all positions of one specific point within a certain time interval [t0; t0 + τ] creates a polynomial curve denoted as path line. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 6
  • 7. Lagrangian Descriptors - 2 ▶ Technique to extract LCS: Finite-Time Lyapunov Exponents (FTLE) which is a method that  describes the motion boundaries, and  separates regions of coherent flow behavior. ▶ Given the optical flow field, FTLE is computed as follows: MMSP’2012 Human Action Recognition using Lagrangian Descriptors 7
  • 8. Lagrangian Descriptors - 3 ▶ Similar motion behavior can be formulated using further geometric features of the path lines. ▶ The areas of similar motion behaviors over time are represented with time-normalized arc length (Ʌ ). arcL ▶ The time-normalized arc length of a given time t0 and an integration interval τ is defined as: MMSP’2012 Human Action Recognition using Lagrangian Descriptors 8
  • 9. Lagrangian Descriptors - 4 MMSP’2012 Human Action Recognition using Lagrangian Descriptors 9
  • 10. Lagrangian Descriptors - 5 ▶ We compute Histogram of oriented Gradients (HoG) to avoid explicit extraction of ridge structures. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 10
  • 11. Lagrangian Classifier ▶ FTLE-HoG and Ʌ -HoG descriptors of video frames provide arcL complementary information about an action. ▶ These two descriptors are fused before training linear multi- class SVMs. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 11
  • 12. Performance Evaluation - 1 Results on the Weizmann dataset Results on the KTH dataset MMSP’2012 Human Action Recognition using Lagrangian Descriptors 12
  • 13. Performance Evaluation - 2 Confusion Matrix on the Weizmann dataset Most confused action pair is skip-jump pair. Confusion Matrix on the KTH dataset Most confused action pairs are • jogging-running, • jogging-walking, • handclapping-hand waving, • boxing-hand clapping. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 13
  • 14. Privacy Aspect ▶ Working with image data introduces another important aspect which is the privacy of the recorded individuals. ▶ In our method, we do not store image sequences directly, but optical flow fields. ▶ Optical flow fields only encode abstract motion patterns. ▶ The recognition of person-related features extracted from abstract Lagrangian motion patterns is much more difficult. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 14
  • 15. Conclusions & Future Work ▶ This study was a starting point to represent & recognize complex human actions. ▶ The main difference from previous works is that our method transforms the motion information of an individual in a given time interval into a 2D space. ▶ It is shown that FTLE and Ʌ Lagrangian measures are well arcL adapted to model individual human activities. ▶ Future Work  Local motion representation – Interest points-based method.  Multi-level motion representation with multi-level Bag-of-Words approach.  Recognizing more complex human motions. MMSP’2012 Human Action Recognition using Lagrangian Descriptors 15
  • 16. Thanks! Questions? MMSP’2012 Human Action Recognition using Lagrangian Descriptors 16