Estimation of Steering Angle and Collision Avoidance for Automated Driving Using Deep Mixture of Experts

Abstract

In this paper, a monocular camera-based method is proposed to estimate the steering angle in autonomous driving. A second-order particle filtering algorithm is used to estimate the steering angles. The filtering algorithm is modeled at the scene-level for varying driving patterns. For a given road scene, individual proposal and likelihood distributions are modeled with deep learning-based regression frameworks for normal driving and obstacle avoidance driving patterns, respectively, the proposal distribution is modeled using a novel long short-term memory-based mixture-of-expert; and the likelihood is modeled using a convolutional neural network. To estimate the driving pattern captured from the monocular camera, a long recurrent convolutional network is adopted and trained. By modeling the distribution at the scene-level for different driving patterns, we accurately model the particle filter distributions. Consequently, for autonomous driving, the steering angle is robustly estimated with few particles. The proposed framework is validated on multiple acquired sequences. A detailed comparative and parametric analysis of the algorithm is performed. The experimental results demonstrate the robustness and accuracy of our filtering algorithm for varying road scenes and driving behaviors.