Event Title

Motion Prediction for Vehicles - CS 4732

Streaming Media

Event Website

https://documentcloud.adobe.com/link/track?uri=urn:aaid:scds:US:ed7d4e8f-3f42-4558-993d-1929dd74ba3a

Document Type

Event

Start Date

3-12-2020 5:00 PM

Description

The idea of a self-driving car is one that makes us think more of a science fiction movie than a plausible addition to real life. While we have not come close to perfecting street legal autonomous vehicles by any means, over the past few years significant progress has been made. Autonomous vehicles need to be able to constantly perceive what is happening in their environment and be able to react accordingly. This study aims to improve upon the accuracy of previous motion prediction models, thus allowing the autonomous vehicle to better understand its surroundings. Building upon Lyft's baseline model, we will implement a model with multi-modal prediction that will generate possible paths that are likely to occur. After training the new model and reviewing the results, the multi-modal approach handles complex intersections well. The baseline model had some instances where a predicted path would ignore the geometry of the road or would cut across an intersection. With the new model, ResNet50 was also implemented. Being a form of deep residual learning, ResNet50 helps give context awareness to the model. The implementation of ResNet50 helped eliminate errors where the predicted path did not follow the geometry of the road, as well as predictions in intersections.

C-Day-Presentation-CS UR-53.mp4 (275832 kB)
Presentation

Share

COinS
 
Dec 3rd, 5:00 PM

Motion Prediction for Vehicles - CS 4732

The idea of a self-driving car is one that makes us think more of a science fiction movie than a plausible addition to real life. While we have not come close to perfecting street legal autonomous vehicles by any means, over the past few years significant progress has been made. Autonomous vehicles need to be able to constantly perceive what is happening in their environment and be able to react accordingly. This study aims to improve upon the accuracy of previous motion prediction models, thus allowing the autonomous vehicle to better understand its surroundings. Building upon Lyft's baseline model, we will implement a model with multi-modal prediction that will generate possible paths that are likely to occur. After training the new model and reviewing the results, the multi-modal approach handles complex intersections well. The baseline model had some instances where a predicted path would ignore the geometry of the road or would cut across an intersection. With the new model, ResNet50 was also implemented. Being a form of deep residual learning, ResNet50 helps give context awareness to the model. The implementation of ResNet50 helped eliminate errors where the predicted path did not follow the geometry of the road, as well as predictions in intersections.

https://digitalcommons.kennesaw.edu/cday/Fall/undergraduateresearch/2