An Improved Approach to 6D Object Pose Tracking in Fast Motion Scenarios

Abstract

Tracking 6D poses of objects in video sequences is important for many applications such as robot manipulation and augmented reality. End-to-end deep learning based 6D pose tracking methods have achieved notable performance both in terms of accuracy and speed on standard benchmarks characterized by slowly varying poses. However, these methods fail to address a key challenge for using 6D pose trackers in fast motion scenarios. The performance of temporal trackers degrades significantly in fast motion scenarios and tracking failures occur frequently. In this work, we propose a framework to make endto-end 6D pose trackers work better for fast motion scenarios. We integrate the "Relative Pose Estimation Network" from an end-to-end 6D pose tracker into an EKF framework. The EKF adopts a constant velocity motion model and its measurement is computed from the output of the "Relative Pose Estimation Network". The proposed method is evaluated on challenging hand-object interaction sequences from the Laval dataset and compared against the original end-to-end pose tracker, referred to as the baseline. Experiments show that integration with EKF significantly improves the tracking performance, achieving a pose detection rate of 85.23% compared to 61.32% achieved by the baseline. The proposed framework exceeds the real-time performance requirement of 30 fps.

DeepKalPose: An enhanced deep‐learning Kalman filter for temporally consistent monocular vehicle pose estimation

“…Given the objective of tracking and correcting pose measurements, the state vector is defined as $$\bm{x}_k = [x_k, y_k, z_k, \theta _k]$$, thereby directing the KF to track the vehicle's pose. However, existing model‐based KF methods [12, 13, 17, 18] have difficulties solving the temporal consistency since most time‐series data from the autonomous driving dataset typically depict vehicles approaching or receding from the camera. Thus, the precision of pose estimation is significantly influenced by the vehicle's position within the image.…”

DeepKalPose: An Enhanced Deep-Learning Kalman Filter for Temporally Consistent Monocular Vehicle Pose Estimation

“…However, existing mathematical model based KF methods [11][12][13][14] have difficulties solving the temporal consistency since most time-series data from the autonomous driving task dataset typically depict vehicles approaching or receding from the camera's viewpoint. Thus, the precision of pose estimation is significantly influenced by the vehicle's position within the image.…”