SteeringNET - A Deep-Learning based Obstacle Avoidance Approach for Autonomous Driving

In this paper, a deep-learning based model named SteeringNET is proposed for obstacles avoidance. Three categories of obstacles were chosen, namely, two-wheelers, cars, and buses. The Ego Vehicle (EGV) was defined using a kinematic model and integrated with the Adaptive Model Predictive Control (MPC) to simulate the obstacle avoidance strategies for human drivers. The dataset for the human driver obstacle avoidance trajectories was generated using MATLAB simulations. A total of 14040 trajectories: 5400, 2160 and 6480 for stationary two-wheeler, car, bus obstacles were obtained. Conventional MPC based models are computationally expensive for obstacle-free trajectory generation and are unsuitable for real-time applications. Our model readily predicts the throttle, steering angle values for negotiating different obstacles and the prediction time is significantly lesser than the Adaptive MPC. Four different versions of this model are implemented namely, Fully Connected Neural Network (FCNN), FCNN with custom loss, Long Short-Term Memory (LSTM) and LSTM with custom loss. The experimental results show that LSTM with custom loss algorithm outperforms the other three deep-learning models.