Methods for remote stock monitoring using depth sensors
RGB-D cameras return images like an ordinary camera but in addition to color, depthmaps where each pixel value represents the distance to a point of the scene are also obtained. Although originally conceived for gaming and consumer applications, their affordably and extensive documentation make them...
| Author: | |
|---|---|
| Format: | doctoral thesis |
| Status: | Published version |
| Publication Date: | 2024 |
| Country: | España |
| Institution: | CBUC, CESCA |
| Repository: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/690602 |
| Online Access: | http://hdl.handle.net/10803/690602 |
| Access Level: | Open access |
| Keyword: | Sitja Silo Volum Volumen Volume Monitoratge Monitorización Monitoring RGB-D Temperatura Temperature Profunditat Profundidad Depth Calibratge Calibración Calibration Càmeres RGB-D Cámaras RGB-D RGB-D cameras 004 62 631 |
| Summary: | RGB-D cameras return images like an ordinary camera but in addition to color, depthmaps where each pixel value represents the distance to a point of the scene are also obtained. Although originally conceived for gaming and consumer applications, their affordably and extensive documentation make them a suitable option for other 3D measurement applications. In this thesis, our interest has been focused on their use for the control of agricultural silo content which is fundamental for its proper management. Despite their potential benefits, obtaining accurate stock data from RGB-D sensors still requires further research and development. In this thesis, three main focuses of research will be considered. • Measurement devices are usually affected by temperature. Centered on RGBD cameras, the first objective is to understand the temperature drift on structured light sensors, characterize it, and propose a compensation model that is fast and reliable to mitigate their effects on the measurements. To reach this objective the sensor performance has been analyzed under different temperature conditions and the distortion model has been characterized as a hyperbolic paraboloid function. We have also proposed a compensation method that reduces the measurement error to the levels of other non-structured light sensors and proprietary solutions. The good results of the method have been demonstrated in real scenarios. • The RGB-D camera position and orientation from devices installed on agricultural silos is crucial for accurate volume estimations. This information is not always available and it is very time-consuming to obtain it by manual inspection. To tackle this problem, a method that uses the shape tensor properties to automatically compute the silo’s axis and provide a new reference system from which the sensor raw data can be easily processed has been proposed. The method has been implemented and tested on both synthetic and real silos, achieving a maximum average distance error of less than 6cm. • Active stereo cameras are commonly designed for close-range applications with medium to low accuracy requirements. However, the standard performance of these cameras, as provided by the manufacturer, may not meet the stringent demands of remote silo monitoring tasks. Especially in cases where content measurements are required for very large silos with high accuracy expectations. To overcome this limitation, a custom calibration pipeline to optimize the calibration parameters of stereo cameras and improve the depth accuracy over long-range measurements has been proposed. The method has been evaluated on a real scenario providing an average relative volumetric error reduction of 8.6% compared to the factory calibration. All the methods presented have been integrated into different production stages of the INSYLO SL services. With more than 500 devices using the thermal compensation algorithms, more than 1000 sensors adjusted using automatic methods, and over 200 sensors calibrated by using the new pipeline. As a result, a substantial increase in the reliability and accuracy of the INSYLO SL sensor network has been achieved |
|---|