Autonomous Robot Exploration with Selective Object Discrimination by Using Deep Learning Object Detection
Blom-Dahl Casanova, Christen (2019)
2019
Science and Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2019-05-31
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201906111880
https://urn.fi/URN:NBN:fi:tty-201906111880
Tiivistelmä
Since the geopolitical world is not polarized anymore, the market competitivity is increasing as never before so in order to survive as an industrial organization, it is key to be competitive. That is, reducing costs and production times among other needs. Mobile robots are resources that manage to get those needs relieved since they can substitute humans and perform better. This causes human issues casuistic drop, human resources re-allocation in more creative job positions which cannot replaced by robots, and more long-term efficiency.
The state-of-the-art of the use of mobile robots remains on the fact that we are talking about not just a single mobile robot but a fleet of them which performs in a smart and coordinated way. These devices can be integrated in the supply-chain so that can transport payloads without the need of any human intervention. In addition, such integration allows a huge flexibility since smart industrial mobile robots can adapt to new conditions, imposed parameters and obstacles that were not predicted. For any autonomous mobile robot, a prior knowledge about its environment is necessary before performing autonomous navigation, that is to have a previous map. Mapping usually is a human intervened task which takes time, especially for large facilities. This work proposes a way to map autonomously, in the most efficient way, an indoor 2D environment by using the Rapidly-exploring Random Trees approach since it is biased towards unknown regions.
In addition, this work proposes object discrimination during mapping. With the conventional approach, during the mapping process laser scanners read the presence of all the obstacles in the environment. This fact is undesired since some of such scanned obstacles are scanned just by causality during the exploration (e.g. personnel, industrial mobile equipment…). Such undesired registered data in the map suppose noise and does not represent the actual long-term environment. The implementation of removing such noise is managed by the combination of two modules. On one hand, by using state-of-the-art deep learning tools in order to achieve real-time object detection. On the other hand, a filter to the laser scanner so that it is blind towards such detections during the exploration, so they are never registered on the map.
The results show quite potential high-quality results which are intrinsically associated with the object detector module. Since such module is state-of-the-art, the technology involved is constantly developing and improving not just the performance but also flexibility and capabilities. This work is a potential new high-fidelity approach besides the conventional approach in order to perform mobile robot exploration.
The state-of-the-art of the use of mobile robots remains on the fact that we are talking about not just a single mobile robot but a fleet of them which performs in a smart and coordinated way. These devices can be integrated in the supply-chain so that can transport payloads without the need of any human intervention. In addition, such integration allows a huge flexibility since smart industrial mobile robots can adapt to new conditions, imposed parameters and obstacles that were not predicted. For any autonomous mobile robot, a prior knowledge about its environment is necessary before performing autonomous navigation, that is to have a previous map. Mapping usually is a human intervened task which takes time, especially for large facilities. This work proposes a way to map autonomously, in the most efficient way, an indoor 2D environment by using the Rapidly-exploring Random Trees approach since it is biased towards unknown regions.
In addition, this work proposes object discrimination during mapping. With the conventional approach, during the mapping process laser scanners read the presence of all the obstacles in the environment. This fact is undesired since some of such scanned obstacles are scanned just by causality during the exploration (e.g. personnel, industrial mobile equipment…). Such undesired registered data in the map suppose noise and does not represent the actual long-term environment. The implementation of removing such noise is managed by the combination of two modules. On one hand, by using state-of-the-art deep learning tools in order to achieve real-time object detection. On the other hand, a filter to the laser scanner so that it is blind towards such detections during the exploration, so they are never registered on the map.
The results show quite potential high-quality results which are intrinsically associated with the object detector module. Since such module is state-of-the-art, the technology involved is constantly developing and improving not just the performance but also flexibility and capabilities. This work is a potential new high-fidelity approach besides the conventional approach in order to perform mobile robot exploration.