Influence of traffic control in a signalized intersection on the risk of road users; Stream-based learning of safety indicators through data selection

Short abstract
In the context of traffic light control strategies and the development of integrated devices for traffic management, our task is to build a system that analyses observation data of an intersection. We use an automated experimental observation device of a real intersection, based on video sensors, and a database of traffic recordings which allows us to make a comparative analysis of the effects of two strategies, a reference strategy, and an adaptive real-time strategy, called CRONOS, developed in INRETS.

We focus on interactions between vehicles, and their relation to accidents, called severity. We introduce a categorization of interactions between vehicles in an intersection, and study particularly the conflict zone. We measure the interaction duration and evaluate at each instant the severity of detected interactions, according to two indicators, a proximity indicator and a speed indicator. Our modular system detects interactions with explicit expert rules. For the severity indicators, we develop a generic learning method, through data selection in a stream. This method is tested on the learning of the speed indicator, and benchmarks. 

We use our system to analyse a part of the database. With this system, we study the distribution of interaction duration according to the safety indicators and highlight differences between the compared strategies. 

Keywords: traffic safety, interactions, safety indicators, risk exposure, machine learning, data selection, incremental algorithms, ensemble methods.