Understanding a dynamic world: Object tracking and dynamic Motion estimation from LIDAR for autonomous driving

Summary

Abstract

In a society that is heavily reliant on personal transportation, autonomous vehicles present an increasingly intriguing technology. They have the potential to save lives, pro- mote efficiency, and enable mobility. However, before this vision becomes a reality, there are a number of challenges that must be solved. One key challenge involves problems in dynamic motion estimation, as it is critical for an autonomous vehicle to have an understanding of the dynamics in its environment for it to operate safely on the road. Accordingly, this thesis presents several algorithms for dynamic motion estimation for autonomous vehicles. We focus on methods using light detection and ranging (LIDAR), a prevalent sensing modality used by autonomous vehicle platforms, due to its advantages over other sensors, such as cameras, including lighting invariance and fidelity of 3D geometric data.

First, we propose a dynamic object tracking algorithm. The proposed method takes as input a stream of LIDAR data from a moving object collected by a multi-sensor platform. It generates an estimate of its trajectory over time and a point cloud model of its shape. We formulate the problem similarly to simultaneous localization and mapping (SLAM), allowing us to leverage existing techniques. Unlike prior work, we properly handle a stream of sensor measurements observed over time by deriving our algorithm using a continuous- time estimation framework. We evaluate our proposed method on a real-world dataset that we collect.

Second, we present a method for scene flow estimation from a stream of LIDAR data. Inspired by optical flow and scene flow from the computer vision community, our framework can estimate dynamic motion in the scene without relying on segmentation and data asso- ciation while still rivaling the results of state-of-the-art object tracking methods. We design our algorithms to exploit a graphics processing unit (GPU), enabling real-time performance.

Third, we leverage deep learning tools to build a feature learning framework that allows us to train an encoding network to estimate features from a LIDAR occupancy grid. The learned feature space describes the geometric and semantic structure of any location observed by the LIDAR data. We formulate the training process so that distances in this learned feature space are meaningful in comparing the similarity of different locations. Accordingly, we demonstrate that using this feature space improves our estimate of the dynamic motion in the environment over time.

In summary, this thesis presents three methods to aid in understanding a dynamic world for autonomous vehicle applications with LIDAR. These methods include a novel object tracking algorithm, a real-time scene flow estimation method, and a feature learning frame- work to aid in dynamic motion estimation. Furthermore, we demonstrate the performance of all our proposed methods on a collection of real-world datasets.

Bibtex entry