Heterogeneous multi-sensor fusion for extended objects in automotive scenarios using Gaussian processes and a GMPHD-filter

Modern advanced driver assistance systems (ADAS) and automated driving functions for automobiles rely on an accurate model of the environment. To this end, the exploitation of complementary advantages of the measurement principles used by radar, lidar and camera sensors is an important prerequisite. We develop a framework for sensor data fusion that incorporates heterogeneous sensor data from multiple sensors in a modular way. In order to use as much genuine information as possible, the measurement data is utilized in its raw low level form or abstracted to a single frame feature representation. Historically, an approach with local decentralized tracks obtained from the individual sensors prevails. This method does not offer true multiple hypothesis considerations in a straight-forward manner. Additionally, a mathematically correct sensor data fusion in the high level approach is infeasible when the covariances of the local tracks are not transmitted. In contrast to this, the full and uncorrelated information contained in the individual measurements provides the possibility for a correct fusion of data and enables a probabilistic conflict resolution of the data association problem. With regard to the multiple hypothesis problem, the Gaussian mixture probability hypothesis density (GMPHD) filter provides a solution. For the estimation of the extent of the observed objects, Gaussian processes offer the possibility to model shapes with a considerable amount of flexibility by using functions which represent the contour of the objects. To demonstrate first results of our approach, we show results with real experimental data from one laser scanner and four short range radars.