Track fusion with incomplete information for automotive smart sensor systems
Ting Yuan, Bharanidhar Duraisamy, Tilo Schwarz, Martin Fritzsche · 2016
Autonomous driving poses unique challenges for vehicle environment perception due to the complex surrounding environment of random and dynamic nature. An autonomous vehicle uses a variety of sensors such as radars, cameras and lidars to obtain the reliable and accurate information on the surrounding environment using a sensor fusion procedure. Each sensor processes data using a local tracker and a track fusion step is carried out to overcome limitations of the individual sensors and to yield more accurate estimation results. However, in a vehicle distributed smart sensor system1, one can get only object tracks but not the corresponding covariance information, primarily due to the safety critical communication channel bandwidth constraint, due to limited computational resources and due to ease of integration. The missing covariance matrix information is critical for a feasible track fusion using classic techniques (e.g., linear minimum mean square error fuser). Moreover, the inner design of each local tracker is unknown to us. This leads to the question on the “credibility” of local object estimates delivered by the sensor and without covariance information the state estimates cannot be properly justified. This leads to a situation where an inconsistent filter might be used and bias (say, due to model mismatch or ego-vehicle pose alignment errors) might be contained in the estimates. Track fusion using incomplete information (i.e., with no covariance/inner design information) is a challenging task. One intuitive approach is to use Kalman filter (KF) directly on the (temporarily autocorrelated) estimates from the smart sensor systems; however, this is statistically forbidden due to violation of input independence assumption (i.e., measurement whiteness) in KF estimation recursions. In this paper, we "imitate" the inner filter design of local trackers and present a de-autocorrelation procedure for the fusion of tracks from each local sensor, and then carry out a filtering stage yielding fused estimate and its associated covariance matrix.