![](https://csdnimg.cn/release/download_crawler_static/89173767/bg4.jpg)
Telecom 2023, 4 462
3. AI, DL and Computer Vision in UAV Sensing
3.1. Artificial Intelligence
The goal of artificial intelligence, a rapidly developing field, is to enable machines to
carry out tasks that would typically require human intelligence, like those made possible
by new information and technological communication systems like the Internet of Things
(IoT). Teaching machines to mimic human behavior is the field’s main goal. Key ideas in
AI include machine learning and deep learning, which entail the construction of multiple
artificial neural network layers. Numerous industries have advanced significantly as a
result of these methods [19–21].
3.2. Deep Learning and Computer Vision
Computer vision is the field of artificial intelligence that enables computer systems to
extract important information from visual inputs such as images and videos and respond
accordingly. Machine vision uses AI to allow computers to perceive, observe, and under-
stand their surroundings. Image processing is used to automate tasks that the human visual
system can perform. To work properly, computer vision requires a large amount of data, as
it analyzes these data to identify and recognize specific inputs, such as frames or images.
To instruct a computer on identifying automobiles, it requires a substantial amount of car
illustrations to enable it to acquire the ability to distinguish and discern cars, including
those that are free from any physical impairment or blemish [
22
]. Neural networks (NNs)
are machine learning techniques that are influenced by biological neural networks. NNs use
nodes (also known as neurons) to transmit information in the same way as human bodies
do. The most frequent type of NN in UAV technology is the convolutional neural network
(CNN). CNNs are a particular kind of NN that is designed for the exclusive purpose of
identifying images. Convolution, as the title suggests, is the technique through which the
initial image at the input of a machine vision application is altered with filters that identify
significant image features like boundaries (Figure 2). The system shall acquire knowledge
about the screened data it identifies independently and correspond it with the anticipated
result. A distinguishable instance is identifying the label of an entity in a specific picture
that serves as an input. Typically, filters consisting of 3
×
3 or
5 × 5 squares
are employed
to detect the orientation of the boundary or characteristic: leftward, rightward, upward,
or downward.
Deep artificial neural networks frequently acquire more efficient data representations
than shallow ones. Nevertheless, incorporating many layers in a CNN can lead to the
problem of vanishing or exploding gradients, which can make optimization challenging [
23
].
In reality, the back-propagation technique is utilized to comprehend the significance of the
filter parameters’ weights utilized in the convolution process.
Figure 2. Training process of a CNN [24].
The utilization of an activation function is a crucial step in the convolution procedure.
Within a neural network, activation functions determine how the input’s weighted sum is
transformed into an output from a node or group of nodes in a specific layer of the network.