Research

I developed an estimation framework for calibrating line-scanning FMCW (frequency-modulated continuous-wave) LiDAR sensors scanning objects on dynamic platforms such as rotating stages and conveyor belts. The framework formulates coordinate-transformation parameter estimation as an optimization problem, exploiting geometric consistency in the reconstructed point cloud as the core criterion. A two-stage solver achieves robust convergence, and post-calibration analysis demonstrates that the recovered point clouds faithfully represent the scanned geometry with a dimensioning error below 0.3%.

I integrated visual perception into a soft robotic manipulator by equipping it with an RGB camera and a time-of-flight (ToF) depth sensor. Multi-day observation of strawberry images in HSV color space revealed that the hue channel is the most informative feature for ripeness estimation. Combining hue-based color filtering with smoothed ToF depth measurements, I estimated strawberry size with a mean absolute percentage error (MAPE) of 10%.

Stereo-cameras provide dense 3D information of the surrounding environment, on the order of tens of millions of points per second. I proposed a bird's-eye-view (BEV) sensor model for stereo-cameras that enables 3D detection of small objects at long range. The sensor model output was consumed by a particle filter, demonstrating real-time tracking of a 16 cm tall object at 160 m distance from a moving ego vehicle.

Collaboration between humans and robotic agents in search-and-rescue missions can ensure the safety of first responders while expediting victim search. I proposed a Bayesian fusion strategy to combine camera-based danger predictions with language-based descriptors provided by a human operator, producing a fused danger estimate for the scene. This estimate was then used to generate a risk-aware search plan for victim localization. A companion dataset was developed comprising realistic search-and-rescue scenes sourced from real-world film footage, labelled with danger-related language attributes.

Appearance attributes have been widely used to identify a person of interest within a gallery of individuals. However, these methods often fail in real-world applications due to incorrect or insufficient attributes. To address the problem of incorrect attributes, I first conducted a human-subject study to identify the appearance attributes most easily recalled by human witnesses, and used them as semantic features for training a zero-shot learning (ZSL) model. To address the complementary problem of insufficient attribute data, I developed an uncertainty-quantification module that iteratively queries the user for additional descriptors about the person of interest when needed.

I studied the consequences of uncertainty in orientation knowledge encountered during the localization and mapping problem in mobile robots. To address that, a two-step approach was proposed where we first estimate the orientation, followed by optimizing for the best position value. A comparison with state-of-the-art methods revealed that this method is tolerant to high sensor noise levels.