Efficient Behaviour based Information Driven Human Tracking System for Long Term Occlusion Recovery

Publication Type:
Thesis
Issue Date:
2024
Full metadata record
Understanding human motion patterns is necessary when developing detection and tracking tools. This work identifies people by analysing torso height and applying classification algorithms. This binary classifier used the Support Vector Machine (SVM) with four features. SVM was selected as the best classifier following extensive experimental testing. The Interacting-Multiple-Model Probabilistic Data Association Filter (IMMPDAF) people tracking method was evaluated using both simulated and experimental data. The evaluation employed NEE and NIS metrics. Robust and consistent IMMPDAF difficulty with targets with long-term occlusion, lowering temporal prediction accuracy. Researchers developed Gaussian processes (GP) to improve tracking during prolonged occlusions. In experiments, the Gaussian Process-Particle Filter (GP-PF) outperformed other methods for predicting time. The computational load increased as more training data samples were added, regardless of accuracy. A sample data management system used mutual information (MI) and Mahalanobis distance to retain useful data while discarding unwanted data. While maintaining the average RMS error (ARMSE) limit, this method significantly reduced sampling data. This research used a 2D laser range finder (LRF) or laser detection and ranging (LiDAR) to observe people indoors. First, select an appropriate detection classification method and integrate it with tracking. Next, a specialised algorithm was developed to improve temporal tracking, particularly in occlusions and partially absent observation data. This thesis combines learning and tracking algorithms to detect and track people using features. The detection and tracking methods made use of learning algorithms as well as parametric and non-parametric regression models. Using laser measurements to group properties for different classifiers made comparing learning algorithms easier, and the confusion matrix assisted in selecting the best detection algorithm. The log-likelihood ratio (LLR) and the Interacting Multiple Model Data Association Filter (IMMPDAF) were used to evaluate how stable and consistent the track generation and termination were. These studies resulted in Gaussian Process-Bayes Filters for long-term occlusion tracking. Furthermore, novel training data management methods reduced training samples while maintaining tracker effectiveness.
Please use this identifier to cite or link to this item: