Drones are being employed in a variety of industries, including agriculture, construction, public safety, and security, among others. Industry analysts anticipate enormous use of these drones in previously unthinkable applications now that deep learning techniques such as computer vision which is powering them.
In this article, we will explore a few of these uses along with difficulties in automating drone-based surveillance.
A real-world example using Nanonets’ machine learning architecture is provided for automated remote inspection of building sites in Africa.
Section 1: Background Information about Aerial Imagery
The general public can now have possession of drones that can fly up to 2 kilometres. These drones are equipped with high-resolution cameras that can take clear pictures that may be utilised for many types of analysis.
Section 2: Industrial drone uses
Aerial images have a variety of industrial uses, including energy (inspecting solar farms), farming (early plant disease detection), public safety (detecting sharks), civil engineering (routine bridge inspections, power line surveillance, and traffic surveying), oil and gas (inspecting on- and offshore oil and gas platforms, drilling rigs), public safeguarding (motor vehicle accidents, nuclear accidents, fundamental fires, ship collisions, plane and train crashes), and others.
Section 3: Obtaining & Editing Industrial-Grade Drone Images
The process of capturing aerial photographs may be summed up in two parts in order to completely capture topography and landscapes.
- Photogammetry: To guarantee that photos overlap throughout a UAV flight, numerous photographs must be collected at regular intervals. This is essential so that measurements may be taken between things visible in the photos. This method is generally referred to as photogrammetry. Relevant metadata must be included during imagery stitching in order for the imagery to be used for data analysis and map creation. A UAV’s onboard microprocessor automatically inserts this metadata.
- Image stitching: Following the completion of data collecting, the next stage is to combine several aerial photos into a useable map. To fast stitch images together, a specialist kind of photogrammetry is usually used. Structure-from-Motion (SfM) photogrammetry is a specialist type of photogrammetry. The SfM programme compares, matches, and measures the angles between items inside each image to stitch together photographs of a single scene taken from various angles. The photographs could be geo-referenced at this stage in order to give each one a location.
Section 4: Drones and Artificial Intelligence
Globally, high-resolution aerial imaging is becoming more and more accessible, and it provides a wealth of data regarding interesting elements that might be connected to upkeep, land development, disease prevention, defect localization, surveillance, etc. Unfortunately, even with intensive manual analysis, it can be difficult to derive useful insights from such highly unstructured data at scale.
Automating the utilisation of aerial photography faces difficulties and limitations.
Automating the processing of drone images presents a number of difficulties. A few of these are listed below with a possible resolution:
- Flat and tiny perspective of Objects: Using human-centric images shot horizontally at close range to the item, current computer vision datasets and algorithms are built and tested in lab settings. The objects of interest for UAV photography captured vertically are often tiny, have fewer characteristics, and mostly seem flat and rectangular. For instance, a UAV photograph of a structure simply depicts the roof, but a terrestrial image of the same building would include elements like doors, windows, and walls.
- Data labelling challenges: Continuing from the previous point, even if we were able to gather a huge number of photos, we would still need to classify them. This is a manual activity that requires accuracy and precision since “garbage in leads to garbage out.” There is no quick fix for labelling that doesn’t involve doing it by hand. Nanonets label the data on demand.
- Large picture sizes: Drone photos are often larger than 3000px by 3000px in resolution. This makes processing such photos more computationally demanding. Nanonets use pre-processing techniques to prepare aerial images for their model training phase in order to get around this. To keep the training invariant to these changes, this entails cropping photographs at various resolutions, angles, and poses.
Section 5: The automation of remote inspection of building projects in Africa
Automating remote inspection of building projects in Africa is covered in Section 5’s Nanonets Case Study.
A South African robotics-as-a-service company named Pragmatic Master worked with Nanonets to automate the remote monitoring of a house building project in Africa.
To track the development of a house in its different stages of construction, Nanonet attempts to find the following infrastructure:
- Foundation (beginning)
- Wallplate (In-progress)
- Roof (Partially finished)
- An Apron (Finishing touches)
- Geyser (Ready to Move In)
sSection 6: Privacy of Data
Trust from Nanonets customers comes first. Their customs are to always provide you full ownership and authority over their content. Two options are available for using their service:
1. Developer: Nanonets may pre-train the models using the customers photos to contribute for their use-case before using them in other applications.
2. Enterprise: Customers own their data! Data will never be used to pre-train any of the algorithms.
Nanonets work alongside their cloud partner, Amazon Web Services, to provide very advanced data privacy and security policies for both plans. As little human involvement as possible is made while pre-processing and training dataset.