Weighted Distance Metrics for Data Association Problem in Multi-Sensor Fusion 2019-01-5022

Traffic accidents are the world's leading threat to human safety. The majority of traffic accidents are due to human error. Advanced Driver Assist Systems (ADAS) can reduce human error, therefore has the potential to effectively improve the safety of road traffic. The perception module in an ADAS understands the surrounding environment of the subject vehicle and therefore is the prerequisite for planning and control. Due to the limitation of computational constrain of Electronic Control Units, ADAS system commonly uses object-leveled multi-sensor fusion, in which raw data is processed to detect objects at the sensory level. In multi-sensor fusion, the task of assigning new observations to the existing tracks, known as Data Association problem, requires distance metrics to present the similarity between tracks. In the literature, metrics, such as standardized Euclidean distance and Mahalanobis distance has been used. Though accounting for the scale and correlation of the data, the existing metrics cannot account for the importance of each feature in predicting their dissimilarity. As a result, weighting factors are added to the distance metrics and they require extensive manual tuning. In this paper, we propose a data-driven method to obtain the weighting factors automatically using supervised learning. Real-world driving data was acquired to train a logistic regression model which obtains the weighting factors based on the predictivity of features. The new distance metric was evaluated using real-world driving data. Comparing to the existing metrics, it achieves better performance in separating dissimilar tracks and higher matching accuracy.