Nachricht

On Wednesday, August 26, Arnold Müller will present the results of his bachelor thesis with the title:

Vehicle Velocity Estimation Using A Monocular RGB Camera

Abstract:

Velocity estimation of automotive vehicles is a challenging problem in the field of Advanced Driver Assistance Systems (ADAS). Traditional methods employ a combination of stereo camera and radar sensors for segmentation and estimation of velocities of other vehicles. However, these sensor-constellations are generally expensive. Therefore, this thesis focuses on an alternative approach using only monocular imagery. The proposed approach is inspired by Kampelmuehler et al. method which involves image feature extraction by an object tracking module and convolutional neural network (CNN) models for depth and optical flow estimation. These features are then processed to serve as inputs to a shallow neural network that regresses the velocity and position of given detected vehicles. Hence, in this work, different combinations of features and parameters for neural networks are evaluated and examined. For training and evaluating the proposed approach, datasets from the TuSimple velocity estimation challenge are used. The results are adequate for the provided dataset. However, how well it performs in general situations is unknown due to the scarcity of available data.