Convolutional Recurrent Neural Network Framework for Autonomous Driving Behavioral Model

Abstract

Autonomous vehicles, without the help of a human, support challenging tasks for sensing the environment and vehicle navigation. The driving behavior is controlled automatically from the observed surroundings using many supervised learning methods that provide action output based on matching the visual inputs and training labels. Most essentially, deep learning algorithms offer improved processing of observed input data but with the increased training, the complexity in processing the real-time data eventually becomes complex. In this paper, an autonomous driving model driven by a behavioral model is designed incorporating (a) recognition, (b) planning and (c) prediction modules. Each module is designed to regulate the processing of input trajectory video data. Additionally, deep learning classifiers are included to improve the automated ability of planning and prediction modules. Initially, the recognition module is planned to limit the redundant data from the raw input data. Secondly, the planning module is designed with convolutional neural network (CNN) to classify the predictable and unpredictable objects from the surrounding trajectories occurring in the line of sight. Finally, the prediction module is designed with recurrent neural network (RNN) to predict the future driving patterns based on the present condition and past driving outputs. The simulation results show that the proposed hybrid deep learning behavioral model offers improved autonomous driving than other existing autonomous driving models. The results of different environments prove that the proposed hybrid model offers increased scalability in terms of improved recall rate of 95.15%, 96.13% and 97.72% in terrain, dense and light traffic zones, respectively, than existing methods.