Examenspresentation av Jingkai Zhou och Qinghuan Liu

Examiner: Lennart Svensson

Abstract:
Advanced Driver Assistance Systems (ADAS) utilizes multiple sensors to help drivers deal with complicated traffic situations. In practice, we need multi-sensor algorithms that can accurately estimate object states, such as positions and velocities, while operating under a limited computational budget.

In this thesis, we propose a multi-sensor fusion algorithm using multi-Bernoulli densities to obtain the fused multi-object densities and estimate the object trajectories. The density of each object is estimated from the single object state fusion, such as covariance intersection, safe fusion, etc. First, we perform data association to associate the densities from different sensors that correspond to the same object. Then, based on the most likely association, the multi-object fusion problem can be decomposed into several single object state fusions. Once the fused multi-object densities are obtained, object trajectories, if desired, can be constructed by backward simulation without object label information. The efficacy of this work has been validated using simulation data as well as real data collected by a Volvo testing truck.

The result shows that our proposed algorithm can accurately associate data in the sense of fewer missed and false detections, achieve accurate multi-sensor multi-object fusion, and the response time is also acceptable.

Welcome!
Jingkai, Qinghuan and Lennart