Collision Avoidance Control for Advanced Driver Assistance System Based on Deep Discriminant Model

Abstract

In this paper, a novel Deep Discriminant Model, DDM is proposed for predicting imminent collisions caused by dangerous lane change, which can be utilized as a collision avoidance control strategy for advanced driver assistance system. Different from previous work, the proposed approach incorporates multiple visual information about the driving environment, as well as the vehicle state and driver's physiological information, information about the uncertainty inherent, and decision making from the spatio-temporal information to the task. In particular, a special network, ConvLSTMs is presented, which is a combination of convolutional and recurrent layers, to process the input image sensor data in both time and spatial domain. The DDM has the ability of extracting features from multiple data sources (e.g., visual, vehicle state and physiological data) in a deep network. Experiments in a simulation environment showed that the DDM can learn to predict impending collisions with an accuracy of 80.8%, especially when multiple modality sensor data are used as input.