|
|
|
We present a monocular visual-inertial odometry (VIO) system that uses only planar features and their induced homographies, during both initialization and sliding-window estimation, for increased robustness and accuracy in dynamic environments. We evaluate on diverse sequences, including our own highly-dynamic simulated dataset, and show significant improvement over a state-of-the-art monocular VIO algorithm in dynamic environments. Project page |
Recent work leveraging learning-based optimizers to tightly-couple correspondence estimation with a weighted least squares objective have shown SOTA results for various pose estimation tasks, but are still difficult to train. We identify possible causes for this instability and propose a simple solution which leads to a 2-2.5x training speedup over a baseline visual odometry model we modify. Coming soon |
Developed an inexpensive safety monitoring system for industrials robots using programmable light curtains, a recently developed controllable depth sensor. The system enables fence-less human-robot collaboration, is flexible and scales easily to many robots, all without compromising on safety. Project page |
We propose a data-driven prior on possible user locations in a map by combining learned spatial map embeddings and temporal odometry embeddings. Our prior learns to encode which map regions are feasible locations for a user more accurately than previous hand-defined methods, and leads to a 49% improvement in inertial-only localization accuracy when used in a particle filter. Project page |
We have developed an autoregressive model to accurately predict future trajectories of traffic participants (vehicles). We demonstrate that using semantics provides a significant boost and allows the model generalize to completely different datasets, collected across several cities, and also across countries where people drive on opposite sides of the road (left-handed vs right-handed driving). Preprint | Video |
We developed a self-supervised deep network, CalibNet, capable of automatically estimating the 6-DoF rigid body transformation between a 3D LiDAR and a 2D camera in real-time. The network alleviates the need for any calibration targets, thereby reducing significant calibration efforts. Preprint | Video |
We developed a suitcase robot that allows blind people to find their way around unfamiliar buildings, by detecting and conveying information about intersections and signs. We conducted a user study with seven blind participants which showed that the robot improved their ability and confidence in navigating compared to their regular aid. Paper | Presentation |
|
Developed a turn-by-turn assistive indoor navigation app for iOS that combined three deep models for localization in real-time -- LSTM for bluetooth-based absolute position estimation, LSTM for IMU-based relative position estimation, LSTM + U-Net for encoding floor map information. Data collected using Meta's Project Aria Glasses were used for training the models. Video demo | Presentation | Meta Connect feature |
|
An end-to-end application with a graphical user interface for easily calibrating the extrinsics between range and visual sensors was developed during GSoC 2018. Automatic and target-less calibration algorithms based on plane-matching and line-matching were integrated into the app, allowing the calibration to be performed in any generic scene setting without the need for any specific targets. Code | Video demo | Report |
|
Mentored and worked with a team in a national-level competition on a janitorial robot to autonomously navigate and clean a washroom setup. The team was selected for the simulation and on-site rounds out of of 136 teams and finished second overall. Challenge page |
|
Developed a visual odometry module based on optic flow for the localization of a custom-built quadcopter and incorporated it into the PX4 navigation stack, enabling autonomous indoor navigation. All the computations were performed on-board, on an Odroid XU4. A stock counting module was implemented using ArUco markers. Report | Code |
|
Ideated and developed an Android application to notify students and faculty about important events, announcements and other campus related information like bus routes and dining menus. It has close to two thousand users today and is the official app of SSN. Code | Store | Appreciation | Press |
|
Built using Hugo and Jon's styling
|