Human-driving highway overtake and its perceived comfort: correlational study using data fusion 2020-01-1036
As an era of autonomous driving approaches, it is necessary to translate handling comfort – currently a responsibility of human drivers – to a vehicle imbedded algorithm. Therefore, it is imperative to understand the relationship between perceived driving comfort and human driving behaviour.
This paper develops a methodology able to generate the information necessary to study how this relationship is expressed in highway overtakes.
To achieve this goal, the approach revolved around the implementation of sensor Data Fusion, by processing data from CAN, camera and LIDAR from experimental tests. A myriad of variables was available, requiring individuating the key-information and parameters for recognition, classification and understanding of the manoeuvres.
The paper presents the methodology and the role each sensor plays, by expanding on three main steps: Data segregation and parameter selection; Manoeuvre detection and processing; Manoeuvre classification and database generation. It also describes the testing setup, and posterior statistical analysis.
To perform all the steps MATLAB was chosen, serving as an all-in-one environment equipped with the necessary toolboxes and libraries to perform filtering, camera perception, operate on matrixes, and database generation.
The resultant algorithms can extract manoeuvres, identify their subsegments (e.g. cut-out, cut-in) and isolate their contribution on the vehicle dynamics and comfort, such as supplemental lateral acceleration. Furthermore, they allow the comparison of different manoeuvres, by grouping them into scenarios conceived through an altered decision tree, in which the selection criteria assimilated a human driver’s decision-making process.
This methodology proved effective on extracting over 300 manoeuvres from 14 experiments, calculating all relative parameters, classifying them into statistically generated scenarios and originating a database useful for statistical analysis, machine learning and manoeuvre generation. In future development, increasing the amount of experiments and diversifying the vehicle types can create a more complete database and a more robust analysis shall be achieved.
Massimiliana Carello, Alessandro Ferraris, Henrique de Carvalho Pinheiro, Diego Cruz Stanke, Giovanni Gabiati, Isabella Camuffo, Massimo Grillo
Politecnico di Torino, Fiat Research Center