Predicting short-term driving intention using recurrent neural network on sequential data

Description

While most of the time, human drivers can predict the simple intentions of other drivers and various on-road behaviors a few seconds in advance, thus rationalizing the associated risks, such reasoning capabilities can be challenging and difficult for an autonomous driving system. Predicting driver intention and behavior is of great importance for the systems that implement safe, defensive path planning and decision making for autonomous driving vehicles. In particular, short-term driving intentions are the fundamental building blocks of relatively long-term and more sophisticated goals, such as overtaking a slow vehicle in front, taking an exit, or merging onto a congested highway.

Zhou Xing shares a methodology that can be used to build and train a predictive driver system, which includes components such as traffic data, a traffic scene generator, a simulation and experimentation platform, a supervised learning framework for sequential data using recurrent neural network (RNN) approach, and validation of the modeling using both quantitative and qualitative methods. The simulation environment can parameterize and configure relatively challenging traffic scenes, customize different vehicle physics and controls for various types of vehicles including cars, SUVs, and trucks, test and utilize a high-definition map of the road model in algorithms, and generate sensor data out of light detection and ranging (lidar), optical wavelength cameras for training deep neural networks and is crucial for driving intention, behavior, and collision risk modeling, since collecting a statistically significant amount of such data as well as experimentation processes in the real world can be extremely time and resource consuming. Standardizing such a testing, scoring system can be very useful to validate and experiment various planning and prediction algorithms of autonomous driving application.