STPOTR: Simultaneous Human Trajectory & Pose Prediction
Non-autoregressive transformer that decodes future 3D pose and global trajectory in parallel for robot follow-ahead, published at ICRA 2023. Cited by 51.
Taher Ahmadi · open-source research and engineering work
A selection spanning computer vision, robotics, and applied machine learning, from my GitHub.
Non-autoregressive transformer that decodes future 3D pose and global trajectory in parallel for robot follow-ahead, published at ICRA 2023. Cited by 51.
Models pedestrians as noisily rational planners, combining probabilistic inference with optimal control to predict where a person is heading in real time, published at ICRA 2022. Cited by 23.
Multimodal dataset of human trajectories, poses, and gaze recorded in an instrumented retail environment, supporting research on socially aware robot navigation. Data in Brief, 2020. Cited by 10.
Added a discriminator and structural-similarity loss on top of a state-of-the-art depth estimator; improved performance on NYU-v2 when adversarial training is used.
B.Sc. thesis: a monocular vision driver-assistance system that detects cars and pedestrians, estimates depth, and warns of probable collisions in real time.
ROS and Gazebo simulator of realistic signal propagation between robots, built at AIRLab (Politecnico di Milano) to study how communication models change exploration algorithms.
Two-stage pipeline: detect lying-down bodies, then reconstruct 3D pose and camera via CNN joints and matching-pursuit sparse representation, all from a single 2D image.
Open-source ROS/Gazebo framework for the Rescue Virtual Robot league: SLAM, navigation, human recognition, multi-agent cooperation. Led the team to 2nd at RoboCup 2017.