Use this URL to cite or link to this record in EThOS:
Title: Algorithms for multi-modal human movement and behaviour monitoring
Author: Townsend, J. S.
Awarding Body: Nottingham Trent University
Current Institution: Nottingham Trent University
Date of Award: 2011
Availability of Full Text:
Access from EThOS:
Access from Institution:
This thesis describes investigations into improvements in the field of automated people tracking using multi-modal infrared (IR) and visible image information. The research question posed is; “To what extent can infrared image information be used to improve visible light based human tracking systems?” Automated passive tracking of human subjects is an active research area which has been approached in many ways. Typical approaches include the segmentation of the foreground, the location of humans, model initialisation and subject tracking. Sensor reliability evaluation and fusion methods are also key research areas in multi-modal systems. Shifting illumination and shadows can cause issues with visible images when attempting to extract foreground regions. Images from thermal IR cameras, which use long-wavelength infrared (LWIR) sensors, demonstrate high invariance to illumination. It is shown that thermal IR images often provide superior foreground masks using pixel level statistical extraction techniques in many scenarios. Experiments are performed to determine if cues are present at the data level that may indicate the quality of the sensor as an input. Modality specific measures are proposed as possible indicators of sensor quality (determined by foreground extraction capability). A sensor and application specific method for scene evaluation is proposed, whereby sensor quality is measured at the pixel level. A neuro-fuzzy inference system is trained using the scene quality measures to assess a series of scenes and make a modality decision. Results show a high degree of accuracy in selecting the optimum modality in a number of separate environmental conditions. The use of colour to identify subjects post-occlusion is typical in tracking. Effectiveness is reduced as the subject count increases with a consequent increased likelihood of similarity between subjects. Experiments are proposed to determine whether a specific histogram parameter configuration, capable of discriminating between subjects in multiple environmental conditions, can be established. An exhaustive search approach for establishing an improved histogram configuration is undertaken using a novel evaluation metric, which assesses the separation of results from intra-subject and intersubject histogram comparisons. Multi-modal, multi-dimensional results show that a 2-D Hue and IR configuration provides greater discrimination than either visible or IR configurations. A tracking system is developed to demonstrate that the methods and configurations can be applied holistically in a real situation. The system is evaluated in a variety of scenarios using challenging subject data aimed at establishing the limits of the system’s capabilities. Through addressing the research question, contributions to the field have been made consisting of: demonstrating the use of a trained neuro-fuzzy inference system to evaluate modality attributes, and the establishment of a generalised multi-modal histogram-based similarity measure to assist in re-establishing subject identity postocclusion. The modular nature of these methods has been demonstrated by inclusion in a developed feature-rich tracking system.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available