Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Acceleration of simultaneous localization and mapping (SLAM) algorithms on graphics processing units (GPUs) for unmanned air drones

Felekis Panagiotis

Full record

---

Summary

In order to achieve fully autonomous work in an unknown environment, many robots rely on cameras and vision algorithms to figure out where to place an object, turn a screw, or weld two pieces of metal together. Mobile robots must solve two basic problems: create a map of the environment and position themselves into this map. Simultaneous localization and mapping (SLAM) algorithm can incrementally construct a map of the robot's surrounding environment while estimating the robot's position in the map. Visual SLAM (vSLAM) uses the camera to obtain corresponding two dimensional digital images from the real three-dimensional world. These camera provides images with high resolution, rich colours and textures where we can exploit to create a very rich map. Due to high computational demands of vSLAM, scaled-down versions are used with smaller resolution and less key features, resulting in poor estimations. In this thesis, we propose an accelerated version of vSLAM that uses a GPU. In our version, we use high resolution images which results in more accurate and rich results. Our system operates in NVIDIA Jetson Tx2 embedded module which is suitable for autonomous robots due to low power consumption. In terms of performance results, our system performs almost identical to a full-powered desktop CPU, while consuming 5x less power. We also prove that our system is as much accurate as other SLAM systems, by using a well-established accuracy dataset.

Available Files

Services