Date of Award

2019

Publication Type

Master Thesis

Degree Name

M.Sc.

Department

Computer Science

Supervisor

Xiaobu Yuan

Rights

info:eu-repo/semantics/openAccess

Abstract

While object identification enables autonomous vehicles to detect and recognize objects from real-time images, pose estimation further enhances their capability of navigating in a dynamically changing environment. This thesis proposes an approach which makes use of keypoint features from 3D object models for recognition and pose estimation of dynamic objects in the context of self-driving vehicles. A voting technique is developed to vote out a suitable model from the repository of 3D models that offers the best match with the dynamic objects in the input image. The matching is done based on the identified keypoints on the image and the keypoints corresponding to each template model stored in the repository. A confidence score value is then assigned to measure the confidence with which the system can confirm the presence of the matched object in the input image. Being dynamic objects with complex structure, human models in the "COCO-DensePose" dataset, along with the DensePose deep-learning model developed by the Facebook research team, have been adopted and integrated into the system for 3D pose estimation of pedestrians on the road. Additionally, object tracking is performed to find the speed and location details for each of the recognized dynamic objects from consecutive image frames of the input video. This research demonstrates with experimental results that the use of 3D object models enhances the confidence of recognition and pose estimation of dynamic objects in the real-time input image. The 3D pose information of the recognized dynamic objects along with their corresponding speed and location information would help the autonomous navigation system of the self-driving cars to take appropriate navigation decisions, thus ensuring smooth and safe driving.

Share

COinS