Offline object matching and evaluation process for verification of autonomous driving
Florback Johan, Lars Tornberg, Nasser Mohammadiha · 2016
We propose an offline method to match tracked objects from two sensors in complex and real life traffic scenarios, which can be used to build an automatic system for offline sensor verification of advanced driver assistance systems and autonomous driving. Detected objects in each of the sensors are described by a dynamic state vector representing their position, speed, and the physical shape of the object. The proposed method is based on a weighted Euclidean distance between a set of features derived from these descriptors. The performance of the developed method is evaluated using a manually labeled data set, where the obtained overall true positive and negative rates are around 85% and 95%, respectively, while substantially better results were obtained considering specific traffic situations such as highways. Additionally, we propose a new method to statistically evaluate the performance of the matching method by running it on the labeled data set and on a larger unlabeled data set.