Deep learning approach for automated geospatial data collection

Geospatial data collection and mapping are considered to be one of the key tasks for many users of spatial information. Traditionally, data collection and mapping can be done using a variety of methods, such as mobile mapping, remote sensing and conventional survey methods. Each method has its advantages, accuracy, costs and limitations. It is therefore essential to assess the requirements of the project in order to ensure that the relevant data quality is acquired at the lowest possible cost. However, one of the greatest barriers is the availability of digital spatial data and attributes. Often this problem arises because these methods are considered costly and require considerable effort and time. With advancements in technology, such as object recognition through Artificial Intelligence technology, this has led to novel approaches to the extraction of features for a number of applications. Information is expected to be more accurate and readily available in real-time at lower operational and field observation costs. Several research groups have therefore investigated the detection of road objects, e.g. road signs. The main drawback of these works, however, is that none of these studies used low-cost sensors to generate geospatial maps in their studies. In addition, some of these studies are considered expensive and require a considerable amount of time to process the information collected. In this study, I presented a new approach to real-time geospatial data collection and map generation by integrating deep learning and geomatics technologies. The proposed solution runs on a laptop which is connected with a single vision sensor, e.g. camera, receiver to capture photographs or videos, and the location unit e.g. using global navigation satellite system to record the user location (geographic coordinates). For some selected classes, a customized data set and a prototype framework "DeepAutoMapping" have been built. "DeepAutoMapping" was developed on the basis of convolutional neural networks inspired by recent rapid advancements in deep learning literature to detect, locate and recognize four main street objects (trees, street light poles, traffic signs, and palms) based on a defined object detection dataset. The prototype calculates the positioning of the detected object using a geographic coordinate system and then generates a geospatial database including object ID, object name, single photograph or video sequence (based on the type of test), distances, bearings, user and object coordinates. It allows users to verify the results in real time without the need to revisit the site. Various evaluation and test scenarios have been conducted to validate outputs. The findings show that the overall proposed approach is easy to use, provides a high detection accuracy of 88% with 6% false detection and a positioning accuracy of 6.16 m for video streaming and 9.99 m for single photography in the outdoor environment. Compared to the current data collection methods available, the proposed solution can be considered as a pipeline for the fastest and cheapest methods of data survey and geospatial map generation. In addition, a new research area for geospatial data collection using deep learning will be opened up.

Text FK 2020 83 - ir.pdf Download (1MB)

Actions (login required)

View Item