Pedestrian Detection and Pose Estimation

Instructor : Dr. Amitabha Mukerjee

Group V6 : Vimal Sharma, Shikhar Sharma

---

Introduction

Pedestrian Detection and Human Pose Estimation is an active field of research. It finds applications in vehicle automation ,video surveillance and content based image retrieval.Variations in appearances caused by different clothing and different configurations of body parts makes it a challenging task.Previous work on this area was either focussed on strong human part detectors or on powerful body models.We seek to combine the strength of both the approaches [1].

---

Intent

We propose to use the generic model[1] for human detection which allows us to detect pedestrians and to estimate their poses.
[1]Pedestrian detection(column1), upper-body pose estimation(column 2), and full body pose estimation(3 and 4)

---

Approach

In our body model,we divide human body into a set of parts. As an example,for pedestrian detection, we consider 8 different parts - head, torso, upper and lower part of left and right legs and feet. We capture probabilistic constraints on part configurations by "prior" p(L).These constraints are mainly due to kinematic dependencies between body parts.

(left)A Kinematic Prior learned on dataset from [3].(right)Some independent samples

We then use Adaboost classifiers to model the representation of body parts.

Comparison between this approach and [3]

---

Possible Extension:

We shall consider action specific constraints in the prior to track physical body movements in a set of images.

---

Dataset Used

TUD-UprightPeople dataset.

---

Project Page

Link of Paper [1] project page.

---

Papers Referred

[1] Mykhaylo Andriluka, Stefan Roth, and Bernt Schiele [2009] Pictorial structures revisited: People detection and articulated pose estimation IEEE Conference on Computer Vision and Pattern Recognition, 2009.
[2]P.F. Felzenszwalb and D. P. Huttenlocher[2005] Pictorial structures for object recognition. IJCV 2005.
[3] D. Ramanan. Learning to parse images of articulated objects[2006] Neural Information Processing Systems 2006. [4] P. Viola and M. Jones. Robust real-time object detection. IJCV, 2004

---