Velocity Estimation from Monocular Video
Velocity estimation of automotive vehicles is a challenging problem in the field of Advanced Driver Assistance Systems (ADAS). Traditional methods used range sensors to estimate the velocity. However, these kinds of sensors are generally expensive.
Therefore, this thesis focuses on the usage of neural networks for real-time velocity estimation of vehicles using only a monocular camera. The proposed approach is inspired by Kampelmu?hler et al. [KMF18] method which involves a vehicle tracking module together with CNN models for depth estimation and optical flow data. The proposed model analyses the input data from the depth and the optical flow to extract features that can be used for velocity estimation. Afterwards, the problem is treated as a regression problem and a shallow neural network is designed to infer the velocity and position of detected vehicles. Hence, in this work, different deep neural network architectures will be evaluated and an appropriate model will be selected. For training and evaluating the proposed approach, datasets from the TuSimple velocity estimation challenge [tus17] will be used.