In applications such as autonomous driving and human-robot interaction, it is important to understand, infer, and anticipate future behavior of pedestrians. We have been investigating and developing novel computer vision and machine learning methods for pedestrian pose, gait, and trajectory prediction. Notably:
In sensor fusion, the aim is to combine the outputs of several algorithms to boost overall performance in applications such as target detection and scene understanding. Previous supervised fusion algorithms often require accurate and precise training labels for each pixel/data point. However, accurate labels are difficult, expensive, or simply impossible to obtain for real, large-scale …
Read more “Multiple Instance Choquet Integral Sensor Fusion With Uncertain Labels”
Synthetic Aperture Sonar (SAS) imaging systems produce high-resolution seabed imagery useful for underwater target detection and scene understanding. One of the challenges for underwater context identification is that the boundaries between seabed contexts (e.g., sand ripple, sea grass, hard-packed sand) are usually gradual with wide regions of transition. We developed a Multiple Instance Learning (MIL) …
Hyperspectral Unmixing refers to the process of estimating the proportions/abundances of each endmember (pure spectral signature of materials) at every pixel of a hyperspectral image. Due to sensor noise and varying environmental and temporal conditions, the spectral signature of the same material may vary across the scene. This phenomenon is called endmember variability. To address …