Please note: This PhD seminar will be given online
Ershad
Banijamali, PhD
candidate
David
R.
Cheriton
School
of
Computer
Science
In this work we propose a method for action-conditional environment prediction for self-driving cars where the environment is represented in the form of Occupancy Grid Map (OGM). Our motivation is that accurate modelling and prediction of the driving environment can efficiently improve path planning and navigation resulting in safe, comfortable and optimum paths in autonomous driving. Due to the importance of interactions between the objects in the scene, it is important to model and predict the driving environment based on the ego-actions to be able to predict the effect of our actions on other agents decisions and behaviours. We train our model in the framework of conditional varitional autoencoders (CVAEs) to maximize the evidence lower bound (ELBO) of the log-likelihood of a conditional observation distribution. We evaluate our model on OGM sequences from NGSIM and Argoverse dataset. The results show significant improvements of the prediction accuracy using our proposed architectures over the state-of-the-art.