Kamrad Khoshhal

Probabilistic LMA-based classification of human behaviour understanding using Power Spectrum technique

This paper proposes a new approach for the Power Spectrum (PS)-based feature extraction applied to probabilistic Laban Movement Analysis (LMA), for the sake of human behaviour understanding. A Bayesian network is presented to understand human action and behaviour based on 3D spatial data and using the LMA concept which is a known human movement descriptor. We have two steps for the classification process. The first step is estimating LMA parameters which are built to describe human motion situation by using some low level features. Then by having these parameters, it is possible to classify different human actions and behaviours. Here, a sample of using 3D acceleration data of six body parts to obtain some LMA parameters and understand some performed actions by human is shown. A new approach is applied to extract features from a signal data such as acceleration using the PS technique to achieve some of LMA parameters. A number of actions are defined, then a Bayesian network is used in learning and classification process. The experimental results prove that the proposed method is able to classify actions.

Trajectory-based human action segmentation

This paper proposes a sliding window approach, whose length and time shift are dynamically adaptable in order to improve model confidence, speed and segmentation accuracy in human action sequences. Activity recognition is the process of inferring an action class from a set of observations acquired by sensors. We address the temporal segmentation problem of body part trajectories in Cartesian Space in which features are generated using Discrete Fast Fourier Transform (DFFT) and Power Spectrum (PS). We pose this as an entropy minimization problem. Using entropy from the classifier output as a feedback parameter, we continuously adjust the two key parameters in a sliding window approach, to maximize the model confidence at every step. The proposed classifier is a Dynamic Bayesian Network (DBN) model where classes are estimated using Bayesian inference. We compare our approach with our previously developed fixed window method. Experiments show that our method accurately recognizes and segments activities, with improved model confidence and faster convergence times, exhibiting anticipatory capabilities. Our work demonstrates that entropy feedback mitigates variability problems, and our method is applicable in research areas where action segmentation and classification is used. A working demo source code is provided online for academical dissemination purposes, by requesting the authors. HighlightsWe develop a entropy feedback model to adjust sliding window parameters.Independent models have been developed for time shift and window size adjustment.The method is generalizable and works in run-time classification.Our results show an improvement in classification precision.The model allows reducing the delay between classified states and ground truth.

HMM-Based Abnormal Behaviour Detection Using Heterogeneous Sensor Network

This paper proposes a HMM-based approach for detecting abnormal situations in some simulated ATM (Automated Teller Machine) scenarios, by using a network of heterogeneous sensors. The applied sensor network comprises of cameras and microphone arrays. The idea is to use such a sensor network in order to detect the normality or abnormality of the scenes in terms of whether a robbery is happening or not. The normal or abnormal event detection is performed in two stages. Firstly, a set of low-level-features (LLFs) is obtained by applying three different classifiers (what are called here as low-level classifiers) in parallel on the input data. The low-level classifiers are namely Laban Movement Analysis (LMA), crowd and audio analysis. Then the obtained LLFs are fed to a concurrent Hidden Markov Model in order to classify the state of the system (what is called here as high-level classification). The attained experimental results validate the applicability and effectiveness of the using heterogeneous sensor network to detect abnormal events in the security applications.

LMA-Based Human Behaviour Analysis Using HMM

In this paper a new body motion-based Human Behaviour Analysing (HBA) approach is proposed for the sake of events classification. Here, the interesting events are as normal and abnormal behaviours in a Automated Teller Machine (ATM) scenario. The concept of Laban Movement Analysis (LMA), which is a known human movement analysing system, is used in order to define and extract sufficient features. A two-phase probabilistic approach have been applied to model the system’s state. Firstly, a Bayesian network is used to estimate LMA-based human movement parameters. Then the sequence of the obtained LMA parameters are used as the inputs of the second phase. As the second phase, the Hidden Markov Model (HMM), which is a well-known approach to deal with the time-sequential data, is used regarding the context of the ATM scenario. The achieved results prove the eligibility and efficiency of the proposed method for the surveillance applications.

A Novel Framework for Data Registration and Data Fusion in Presence of Multi-modal Sensors

This article presents a novel framework to register and fuse hetero- geneous sensory data. Our approach is based on geometrically registration of sensory data onto a set of virtual parallel planes and then applying an occu- pancy grid for each layer. This framework is useful in surveillance applications in presence of multi-modal sensors and can be used specially in tracking and human behavior understanding areas. The multi-modal sensors set in this work comprises of some cameras, inertial measurement sensors (IMU), laser range finders (LRF) and a binaural sensing system. For registering data from each one of these sensors an individual approach is proposed. After registering multi- modal sensory data on various geometrically parallel planes, a two-dimensional occupancy grid (as a layer) is applied for each plane.