A machine-learning approach to reduce the risk of collision when changing lanes

Abstract

Abstract Advanced driver assistance systems support drivers to handle different complex traffic conditions. The support systems get traffic information from sensors and use algorithms to avoid risks. However, dealing with complex time series data from various sensors is challenging. In this thesis, a machine learning approach is proposed for threat assessment for lane changing. The emphasis is on vehicle state prediction and maneuvers for autonomous emergency steering. The work includes feature selection, model selection, and model validation. Feature selection is performed using the NSGA-II algorithm and correlation analysis to identify the most influencing features. This helps reduce the data dimension while maintaining prediction accuracy. An artificial neural network model structure inspired by ResNet is developed. This network structure is built from blocks, each with a shortcut. Various model configurations, including the number of input features and the network depth, are tested to find a reasonable tradeoff. In addition, driver-state information is also analyzed, and the "most probable gaze zone" data features enhance the model’s performance. The proposed model is validated on real-world data and has good performance.