David Harwood

3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

David Harwood focuses on computer vision, specifically object detection and human tracking in complex visual environments. His trajectory spans the application of partial least squares analysis to high-dimensional feature spaces for identifying humans and vehicles, as well as the use of mixed network search strategies for tracking articulated human motion. The most recent work in the evidence addresses multi-person tracking under occlusion.

Research Trajectory

The available evidence outlines a focused period of research activity between 2009 and 2013, primarily centered on advancing object detection and human tracking methodologies in computer vision. A recurring technical challenge in this work is the management of extremely high-dimensional feature spaces and complex spatial constraints.

In 2009, Harwood participated as a co-author in research addressing human detection. The paper Human detection using partial least squares analysis explored methods to augment standard edge-based features with texture and color information. Because this augmentation produced a feature space exceeding 170,000 dimensions—rendering classical support vector machines computationally intractable—the research utilized partial least squares (PLS) analysis to handle the high dimensionality. This work intersects with the broader research direction of Robust Edge Detection in High Noise Environments, as it sought to build richer descriptor sets beyond fragile local intensity gradients to reliably identify human figures.

Building on the utility of PLS for high-dimensional visual data, subsequent collaborative work applied similar dimensionality reduction techniques to aerial imagery. As a second author on Vehicle Detection Using Partial Least Squares. (2011), Harwood contributed to a vehicle detection framework that integrated Histograms of Oriented Gradients (HOG), a novel "Color Probability Maps" feature, and structural descriptors called "Pairs of Pixels." The combination again resulted in a massive feature set of approximately 70,000 elements, which was efficiently processed using PLS to improve detection accuracy in urban planning and visual surveillance contexts.

By 2013, the focus of the documented research shifted from static object detection to the dynamic tracking of articulated human figures. In Tracking People's Hands and Feet Using Mixed Network AND/OR Search, Harwood and collaborators introduced a framework for tracking the extremities of multiple interacting humans from a single 2D camera view. Rather than relying on multiple views, high-resolution imagery, or large training sets, this approach leveraged mixed probabilistic and deterministic networks. The system employed an AND/OR Branch-and-Bound search with lazy evaluation to resolve the complex interperson occlusions that naturally occur during human interaction.

Current Directions

Based on the provided corpus, the most recent paper in the evidence is from 2013. This later work investigates the tracking of articulated human motion and the resolution of complex visual occlusions. The primary methodological focus in this period involves using mixed network AND/OR search spaces and deterministic bounding techniques to track multiple interacting individuals efficiently, moving beyond static detection toward continuous spatial tracking in constrained environments.

Profile updated

See something wrong? Report a mistake