Probabilistic multi-object tracking for autonomous vehicles
Motro, Michael, 0000-0002-0259-4514 · 2019
Interactive robots such as self-driving cars require accurate hardware and methods to locate relevant objects such as other traffic participants. They also must predict other participants' actions or understand their role in the environment. Given imperfect information about present objects at each time, a multi-object tracker maintains an estimate of all present relevant objects and infers motion or other information that can be deduced from viewing an object over time. Trackers are often built around a probabilistic model that includes known characteristics of object motion and sensor behavior. This thesis discusses several details for designing a probabilistic multi-object tracker for vehicular environments, as well as ways to utilize probabilistic tracked estimates for autonomous vehicle applications. The increasingly complex environments perceived by robots have demanded new paradigms of perception. In particular, camera and laser-based perception of urban settings is solved using learned algorithms that directly transform raw data into object estimates. We present a probabilistic model of modern object detectors that can be integrated with standard trackers. The primary effects that are modelled are line-of-sight limitations to sensor detection, and correlation in algorithmic detection errors over time. Each of these modifications are shown to improve performance on a public benchmark for vehicle tracking, without fundamental modifications to the tracking algorithm. Accurate tracking can require intensive computation on its own. We examine the implementation of multiple hypothesis tracking, a high-performance probabilistic tracker, and improve the computational efficiency of its data association algorithm in several ways. The modified algorithm is tested on vehicular tracking data as well as simulated large-scale and multisensor problems. The improved speed of the algorithm allows for more hypotheses to be propagated at a given speed, which in turn improves tracking performance. In addition to improving the current estimate of the environment, tracking enables prediction of the future environment by determining object motion and history. The uncertainty of these estimates can be quantified by a probabilistic tracker and should be considered when making predictions or deciding actions. However, probabilistic estimates are difficult to translate into interpretable and actionable concepts, such detection of impending collisions between objects. We disambiguate the error rate in collision detection into inevitable errors from uncertain object estimation and further errors incurred by fast approximate calculation of the probability of collision from these estimates. Various methods for collision detection from uncertain data are compared and tested on vehicle simulations. Automated overtaking assistants are studied as a specific application of collision detection. These assistants alert drivers in advance that entering the opposite lane to pass a slower vehicle will be unsafe. We characterize the expected design of these systems, including sensor or communication accuracy and limitations as well as driver variability and uncertainty in future motion. Overtaking assistant simulations demonstrate that the assistant can fulfill its purpose at expected levels of tracking and prediction uncertainty, provided that the chosen sensor or communicating device has a sufficient operating distance