Enhancing point clouds using low end devices

Enhancing point clouds using low end devices

Kreuk, L.I.
van den Berg, R.J.C.
Hammudoglu, J.S.

van Gemert, J.C. (mentor)

Electrical Engineering, Mathematics and Computer Science

Computer science

Technische Informatica

TI3806

2016-06-24

Currently, after a traffic accident happens and the police arrive at the scene it takes some time for them to gather the necessary evidence. This means that the road is closed for quite some time. To speed this up our client Geodelta has developed Raida. The idea is that the police will take photos of the area using a calibrated DSLR camera. They will then use Raida to generate an accurate computer model of the scene. This model can then be used to gather further evidence, such as measurements of tire tracks. This means that less time is needed on the scene, and the road can be reopened sooner. Our software will be using that structure. The pictures and other data from Raida will be referred to as the Raida data. Often, the police are able to gather photos taken by bystanders using low-end uncalibrated devices, such as smart phones. Many of these photos are taken before the police arrive, which mean that they can contain relevant information that is no longer present when the police take their photos. Currently, the police is uncertain on what to do with these photos. It would be useful to be able to use them to enhance the Raida model. To this end Geodelta asked us to create a way to use smartphone photographs to enhance the Raida model. In this report we will describe the software product we have created. Our product is called Visual Pointcloud Enhancer, or VIPE. It will take as input an accurate 3D model in the form of a point cloud, the photos used to generate this point cloud, and a number of additional uncalibrated photos. It will show additional information about the scene. This report will discuss the algorithms and libraries used for developing VIPE. As Raida has a very limited visualisation of the point cloud, VIPE focuses on a clearer and more informative visualisation of the point cloud so the point cloud can also be used to examine the accident scene. The external photographs are linked to the Raida data using a SfM pipeline which produces a 3D point cloud. The point cloud consists of feature points from the pictures and the camera positions of these pictures. We know that the Raida data is relatively accurate as there is a quality control in the program. We calculate a transformation matrix by comparing the coordinates of the Raida data en SfM pipeline data and can add with this transformation matrix the camera positions of the external data to the Raida data. Within the pipeline SIFT is used to extract features from the photographs and match these features using a nearest neighbour algorithm. While implementing the software some important software goals were taken into account. Geodelta wants to use our software and extend it or use parts of it. Documentation and readability of the code was focused on intensively during the development phase. Also modularity and extensibility were taken in account when writing the software. Testing is also an important part of the code. The software is tested thoroughly using unit tests and continuous integration. These unit tests were executed on Linux and Windows to guarantee VIPE works on both platforms. Also a user test is executed, to check if the software program is clear to the user. The software product we made is a working beta implementation that includes all the requirements set with the client. The product can be extended or partly used for future products of Geodelta.

http://resolver.tudelft.nl/uuid:097b7eb7-b7f7-442b-8c5d-151f6bf4ad20

Student theses

bachelor thesis

(c) 2016 The Author(s)