Deep Learning Based Motion Forecasting for Autonomous Driving
2021
Hochschulschrift
Zugriff:
Autonomous driving (AD) is a promising technology that has grown into one of the primary areas of interest for the controls and machine learning research communities and commercial companies in the current decade. Solving this challenging problem can have significant impact on society and in the way we commute and interact with our surrounding environment. Studies have shown that self-driving vehicles could help significantly reduce loss of life due to road accidents and also reduce carbon emissions. Simply put, AD can make travelling from one point to another more efficient and eco-friendly. There's been a recent boom in number of research labs and companies working on this problem which can be accredited to the success of deep neural networks in solving challenging tasks involving learning such as image recognition and language translation. This has driven the applied machine learning community to implement the state of the art techniques in the field to the problem of self-driving. Although, this problem turns out to be much more complicated than the former due to the amount of uncertainty in the dynamic environment and the decision making involved. Nevertheless, there has been a significant progress in the field in the area of object tracking, motion forecasting and path planning and the self-driving pipeline is rapidly evolving.In this thesis, one such module of the self-driving pipeline known as motion forecasting is tackled. Motion forecasting is tasked with predicting the future states of all the agents in a given scene that affect the future trajectory of the self-driving agent. This component takes input from the object detection and tracking module and predicts the future trajectories of all agents that affect the behaviour of the self-driving ego agent. The output from this module helps the ego agent perform path planning efficiently to drive itself from origin to destination. The architecture for the model proposed in the thesis is built on deep learning based neural networks as they are appropriate for this data-driven learning task. After discussing different approaches to solving the problem in hand from the literature survey and highlighting some related work, the thesis focuses mainly on demonstrating two novel proposed architectures based on the encoder-decoder design, discussing the methodology behind the design along with the merits and demerits. The proposed design includes elements such as spatio-temporal encoding to efficiently learn the static map and temporal agent states, agent interaction modeling with different attention mechanisms to capture interdependence between various agent behaviors and specialized basis based trajectory decoders to precisely estimate the future trajectories of the agents. This is followed by a series of experiments and results that demonstrate the incremental improvement made in the model design to achieve state of the art results and the performance of the models on a popular motion prediction dataset. Additionally, a meta-learning based test time adaptation technique is proposed to perform test-time single shot correction of the predictions for newly encountered driving scenarios.Solving the problem of motion forecasting is a major step towards achieving Level 5 self-driving, and the proposed models in this thesis are a small contribution to the progress to be made in solving this challenging tasks.
Titel: |
Deep Learning Based Motion Forecasting for Autonomous Driving
|
---|---|
Autor/in / Beteiligte Person: | Dsouza, Rodney Gracian |
Link: | |
Veröffentlichung: | 2021 |
Medientyp: | Hochschulschrift |
Schlagwort: |
|
Sonstiges: |
|