Publications‎ > ‎

Project Page


Aerial Suspicious Action Detection

Highlights:

The details for the paper "Autonomous UAV for Suspicious Action Detection using Pictorial Human Pose Estimation and Classification" published in Electronic Letters on Computer Vision and Image Analysis, available at: [Link], are provided below: 

Ø  Proposed an autonomous unmanned aerial vehicle (UAV) that locates humans in image frames.

Ø  Pose is estimated as a pictorial structure using weak constraints on position, appearance of body parts and image parsing.

Ø  Suspicious action dataset was created.

Ø  Estimated pose is flagged using Hough Orientation Calculator on close resemblance with any pose in the dataset.  

Ø  The system produces an accuracy of 71%.


Video:





Algorithm:

Figure 1: Processing pipeline for calculating pictorial pose



Qualitative Results:

The proposed algorithm is tested on three different videos recorded in various environments, including different suspicious actions. We applied the proposed algorithm to 170 images containing 430 poses for video-1, 310 images with 873 poses for video 2 and 152 images with 367 poses for video-3 respectively. The images depicted in the paper have been captured using a camera mounted on an UAV. The camera was fixed at 25 degree with respect to the ground in order to capture maximum number of people in the frame. The camera is capable of recording high definition videos at a frame rate of 60fps. The sequence recorded was about two minutes long summing to 7200 frames per sequence. Results of the algorithm on a frame recorded from the UAV is shown below.


Figure 2: Results of the algorithm on a frame recorded from the UAV (a) Result after removing overlapped bounding boxes (b) CRF figure representation of each person (c) Result marking the suspicious person hitting the other person is marked (in red).


Quantitative Results:

Figure 3: The graphs shows (a) Action Recognition accuracy and (b) Average time complexity.


Maltab Code:

Ø The code can be download from this link [Code].

Ø Certain parts of our code are adapted from Upper-body detector available at [Link].


Terms of Use:

If you choose to use our work in your research, please cite the following paper:

@article{ELCVIA582,
	author = {Surya Penmetsa and Fatima Minhuj and Amarjot Singh and SN Omkar},
	title = {Autonomous UAV for Suspicious Action Detection using Pictorial Human Pose Estimation and Classification},
	journal = {Electronic Letters on Computer Vision and Image Analysis},
	volume = {13},
	number = {1},
	year = {2014},}

References:

[1] Eichner, M. and Ferrari, V. Better Appearance Models for Pictorial Structures Proceedings of British Machine Vision Conference (BMVC), 2009. 

[2] M. Marin, V. Ferrari, A. Zisserman upper-body detector www.robots.ox.ac.uk/~vgg/software/UpperBody/

[3] P. Felzenszwalb, R. Girshick, D. McAllester, D. Ramanan Object Detection with Discriminatively Trained Part Based Models Pattern Recognition and Machine Learning (PAMI), 2009. 

[4] P. Viola, M. Jones Rapid Object Detection using a Boosted Cascade of Simple Features Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2001.

[5] N. Dalal and B. Triggs Histograms of Oriented Gradients for Human Detection Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2005.

[6] M.Eichner, M. Marin-Jimenez, A. Zisserman, V.Ferrari Articulated Human Pose Estimation and Search in (Almost) Unconstrained Still Images ETH Zurich, D-ITET, BIWI, Technical Report No.272, September 2010. 

[7] P. Felzenszwalb, D. McAllester, D. Ramanan A Discriminatively Trained, Multiscale, Deformable Part Model IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2008.













Visits




Comments