Authors:
Razvan Itu
and
Radu Danescu
Affiliation:
Computer Science Department, Technical University of Cluj-Napoca, Memorandumului 28, Cluj-Napoca, Romania
Keyword(s):
Autonomous Vehicles, Self-Driving Vehicles, Convolutional Neural Network, Velocity Prediction, Vehicle Speed, Recurrent Neural Network, Long Short-Term Memory, Vehicle State Estimation, Visual Odometry.
Abstract:
Detecting the ego-vehicle state is a challenging problem in the context of autonomous vehicles. Perception-based methods leverage information from on-board cameras and sensors to determine the surrounding traffic scene and vehicle state. Monocular based approaches are becoming more popular for driver assistance, and accurate vehicle speed prediction plays an important role for improving road safety. This research paper presents an implementation of a Convolutional Neural Network (CNN) model for vehicle velocity prediction using sequential image input, as well as an extended model that also features sensorial data as input. The CNN model is trained on a dataset featuring sets of 20 sequential images, captured from a moving car in a road traffic scene. The aim of the model is to predict the current vehicle speed based on the information encoded in the previous 20 frames. The model architecture consists of convolutional layers followed by fully connected layers, having a linear output l
ayer for the ego-vehicle velocity prediction. We evaluate our proposed models and compare them using existing published work that features Recurrent Neural Networks (RNNs). We also examine the prediction of the brake pedal pressure required while driving.
(More)