Efficient Hybrid Machine Learning Algorithm for Trajectory Planning in Critical Traffic-Scenarios

Recently proposed algorithms , namely the Hybrid Augmented CL-RRT, the Hybrid Augmented CL-RRT+ and GATE-ARRT+, reduce the computation time for safe trajectory planning drastically using a combination of a deep learning algorithm 3D-ConvNet with a vehicle dynamic model. In order to realize these algorithms in a vehicle, an efficient embedded-implementation of the algorithms in an automotive micro-controller is required as the on-board micro-controller resources are limited. This paper proposes methodologies for replacing the computationally intensive modules of trajectory planning algorithms such as checking for collisions with traffic participants predictions using machine learning algorithms and analytical methods for reducing the required static RAM memory. After optimising the algorithms, the results are generated by downloading and running the algorithms on various hardware platforms: automotive micro-controller, a rapid prototyping hardware and raspberry pi. The results show that significant reduction in computational resources and potential of proposed algorithms in real time.