BEWARE Project

Human Action Recognition

Introduction

In the proposed work, human action recognition is performed using a simplified human skeletal structure composed by 4 joined points; these points are representing the feet, the abdomen and the head. By means a Hidden Markov Model technique are obtained the recognition results observing how the skeletal structure evolves over the time. This solution appears robust and reliable over different scenario, moreover it provides a method to identify the most common actions generally detected in public area or transport, using a simple and low computational cost framework.

Our approach

To describe the human body shape using a 4 points skeletal structure it is necessary to precisely localize saliency body parts, in our case we select head, feet, and abdomen. Moreover it is needed to describe over the time the structure movements. Principal advance of this idea consists in the opportunity of overcoming different problems like partial abdomen occlusion or strong shadow; this becomes possible forcing a structuremodel updating that allows to recover skeleton’s missing points.

DETECTION:

FRAMEWORK FOR HEAD DETECTION
FRAMEWORK FOR ABDMOMEN DETECTION
FRAMEWORK FOR FEET DETECTION

REGOGNITION: Since the number of parameters required to describe the model is unacceptably high when the feature vector is of high dimension. Learning in high dimensions may lead to data over-fitting therefore generalising poorly on unseen data. To address this problem, a Multi-Observation Hidden Markov Model (MOHMM) is employed to model the action cycles. The models employed in our study is restricted as a left-right model, i.e. no transitions are allowed to states whose indices are lower than the current state. Additional constraint is placed so that no jumps of more than one state are allowed.

Results

We evaluate the proposed framework by using Weizmann action database .  We have chosen five specific classes of actions (a total of 48 sequences) having high interclass similarity. Specifically, the chosen sequences consists actions performed by nine different people. The actions include running, walking, skipping, jumping forward on two legs and gallopping sideway. The obtained recognition rate is the following:

Demos

WALKING.avi LIMP.avi BAG.avi 

Publicatons
Updated November 2008