Proceedings of the 2023 International Conference on Machine Learning and Automation

DOI: 10.54254/2755-2721/32/20230219

YOLO model-based target detection algorithm for UAV

images

Anqi Wei
School of Communication & Information Engineering, Shanghai University, Shanghai,
200444, China

[email protected]

Abstract. The increasing popularity of drones has paved the way for their utilization in various
sectors, including civil, commercial, and government agencies. These unmanned aerial vehicles
have proven to be invaluable in capturing images and videos from vantage points that were
once difficult to access, leading to a wide range of applications. Images captured by drones
often have target objects that are small in the frame and a large number of photos or videos
captured, so that it is difficult for people to find the target objects in the photos. Nowadays,
target detection of images captured by drones through deep learning methods, such as the
YOLO algorithm, can greatly help people's work. In this paper, the authors of this paper have
investigated for the last three years, for target detection of UAV images, optimization based on
the original YOLO algorithm to achieve improved detection results. The research in this paper
summarizes the existing research results and is of great significance to the subsequent research
and application of UAV image processing.

Keywords: yolo, drone image, UAV image, target detection.

1. Introduction
Commercial small aerial vehicles, also known as drones, have the advantage of being portable and
more flexible during flight. Drones on the market today are often equipped with high-definition
cameras and have real-time sharing capabilities that allow users to capture images and analyze them at
any time. Nowadays, using drones for real-time monitoring is becoming increasingly common, and the
technology can be applied to human flow monitoring, road traffic flow assessment, forest fire
inspection, and large motor equipment inspection. The technology reduces the need for personnel for
such tasks or reduces the pressure on personnel, and personnel are less likely to need to travel to
dangerous areas, providing more security for staff. The use of drones for inspection and finding, in
addition to the hardware equipment requirements, also needs to be able to picture quickly to complete
the target detection and the need for relatively high recognition accuracy.
There are the following problems to be solved in target recognition of images using UAVs. The
first one is the dataset. In early image target detection algorithms, such as face recognition, license
plate recognition, etc., most images are portraits or the front of the object. Much research is also based
on the training of such datasets obtained. However, due to the flight characteristics of UAVs, the
images captured are all top-down views, and the algorithms also need to be retrained on the data under
this characteristic. Second, the size of the target detection object captured by the UAV, even at high

© 2023 The Authors. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0
(https://creativecommons.org/licenses/by/4.0/).

248
 Proceedings of the 2023 International Conference on Machine Learning and Automation
DOI: 10.54254/2755-2721/32/20230219

resolution, will be smaller than the target object captured by the flat camera. The algorithm needs to
enhance the detection performance based on this property, such as improving the characterization of
target features, which can improve the correct rate of target object detection. In addition, images
captured by UAVs often have the problem of rotational angle because it is impossible to restrict the
UAV's camera during flight [1]. Therefore, the algorithm's detection cannot be limited to horizontal or
vertical situations but must have good detection for any rotation angle.
As a popular mainstream target detection algorithm, the YOLO algorithm is fast, small, and
guarantees a good detection rate [2]. Many researchers have based on this algorithm framework and
further optimized the framework for the characteristics of UAV images. In this paper, the authors
investigate the research results in this direction in the last three years, so that the subsequent research
can be further deepened.

2. Research results in the past three years

The YOLO algorithm is one of the most popular deep learning-based target detection algorithms and
has now been released in YOLOv7. YOLOv4 builds on YOLOv3, a major breakthrough in real-time
target detection and improved accuracy. YOLOv5 builds on the success of YOLOv4, further
improving accuracy and enhancing the model's generalization capabilities. YOLOv7 is the latest
version, optimized for speed and accuracy.
Luís Augusto Silva et al. used YOLOv4, YOLOv5, and YOLOv7 for target detection in UAV
images, and achieved 59.9% mAP in YOLOv5, 65.70% mAP in YOLOv5 with Transformer
Prediction, and 73.2% mAP in YOLOv7 [3]. For UAV images, Transformer Prediction Head (TPH) is
added to YOLOv5, improving the large-scale change of target objects in UAV images. And effective
methods for target detection in UAV images were screened, and self-trained classifiers were used to
improve the classification accuracy under too vague criteria. The project also improved the
classification of the dataset of pavement damage, which made the algorithm more effective in
detection through more detailed classification.
Jinsu An et al. proposed a method to improve the YOLOv5 network with CBAM to implement an
algorithm for target detection for UAV vision and achieved a mAP of 22.56% [4]. The addition of
CBAM to the network of the original YOLOv5 realized the introduction of combinatorial fast
structures. Because CBAM combines channel attention and spatial attention, the improved YOLOv5
has a convolutional block attention module. CBAM strengthens the performance of the attention
module based on the BAM module while remaining lightweight and generalized, and can be flexibly
added within any CNN architecture and train the model end-to-end. During the optimization of the
three parts of YOLOv5, CSPDarknet53 was used for the backbone, PANet for the neck, and B×(5+C)
for the output layer of the head. With this optimization, the algorithm has a better performance for the
extraction of feature information.
Zhengwei Li et al. proposed R-YOLOv5, a lightweight rotating target detection algorithm, which
achieves an mAP of more than 80% on different datasets and shows good generalizability [5]. The
algorithm is based on YOLOv5 and optimized for the backbone feature extraction network, the neck
feature fusion network, and the prediction head. The angle prediction branch is combined in the part of
the prediction head, and the Circular Smooth Labeling (CSL) angle classification method is introduced
to be able to measure the distance between the angle labels, which makes YOLOv5 able to detect the
scenes with unknown rotation angles. The problem of tiny objects in UAV images not being able to
retain feature information to higher level feature information after multiple convolution operations is
also partially solved by feature fusion embedded in the Swin Transformer block (STrB). In UAV
images, there is also the problem that similar objects cannot be detected correctly due to more noise in
the feature information. Using the Feature Enhanced Attention Module (FEAM), which incorporates
the Multihead Self Attention (MHSA) module, enhances the ability of the network to capture the
information and ensures detection accuracy. There is a problem of object scale distortion in the images
captured by the UAV. By adding the Adaptive Spatial Feature Fusion Structure (ASFF) to the head of
YOLOv5, the algorithm can adapt to objects of different scales and no longer loses object information.

249
 Proceedings of the 2023 International Conference on Machine Learning and Automation
DOI: 10.54254/2755-2721/32/20230219

The algorithm reduces the excessive computational complexity during feature fusion in the backbone
network, increases the utilization of detailed information, and improves multi-scale feature fusion in
the head.
Oyku Sahin et al. In response to the problems of high viewing angles, large target scale variations
and unstable image quality of UAV images, and an improved network structure was used [6]. The
network structure utilizes a combination of Convolutional Neural Networks (CNNs) and Feature
Pyramid Networks (FPNs) to effectively capture target information at different scales in the image,
thus improving target detection accuracy. The team also proposed a new loss function for optimizing
the training process of the network. This loss function combines the position loss and confidence loss
of the target frame, as well as the categorization loss, which integrates multiple aspects of target
detection and enables the network to better learn the position and category information of the target. In
addition, to cope with the problem of changing target scales in UAV images, a target scaling technique
is introduced in the paper for appropriate processing of targets at different scales in the image,
improving the robustness and performance of target detection. The experimental results show that
YOLODrone can achieve higher target detection accuracy and faster detection speed in UAV images,
proving the superiority and practicality of the method.
Sushil Kumar et al. utilized deep learning techniques, based on the YOLO V5 algorithm, to
improve the accuracy and efficiency of target detection in Unmanned Aerial Vehicle (UAV)
surveillance images [7]. The algorithm adopts a one-stage detection strategy, transforms the target
detection problem into a regression problem, and utilizes a feature pyramid network (FPN) for multi-
scale feature fusion, which results in an excellent performance in dealing with target scale variations.
Optimizing the YOLO V5 algorithm for the characteristics of UAV surveillance images, the
researchers introduced a series of improvements. The method can better adapt to the complex scenes
of UAV surveillance images by introducing special convolutional layers, attention mechanisms and
data enhancement techniques. Meanwhile, applying pre-training and migration learning enables the
model to be trained on small sample data with better generalization ability. To verify the method's
effectiveness, the researchers constructed a UAV surveillance image dataset containing various types
of targets and conducted extensive experiments on the dataset. The experimental results show that the
target detection and recognition method based on the YOLO V5 algorithm achieves significant
performance improvement in UAV surveillance images. Compared with traditional methods and
baseline models, the method has significant advantages in target detection accuracy and efficiency,
and is able to identify and localize various types of targets in surveillance scenes more rapidly.
Weibiao Chen et al. proposed the DSM-YOLO v5 algorithm, which aims to improve the accuracy
and efficiency of target detection in UAV aerial images [8]. The paper chose the YOLO v5 algorithm
as the basic framework, and the algorithm adopts a one-stage (one-stage) detection strategy, which
performs well in dealing with target scale variations by transforming the target detection problem into
a regression problem, as well as utilizing a feature pyramid network (FPN) to achieve multi-scale
feature fusion. Optimized for the characteristics of UAV aerial images, the paper proposes the DSM
(Digital Surface Model) mechanism. The DSM technique further improves the accuracy and
robustness of target detection by acquiring the surface elevation information and fusing it into the
YOLO v5 algorithm. The introduction of the DSM helps to better localize and recognize the target.
The experimental results show that the DSM-YOLO v5 algorithm is able to achieve significant
performance improvement in UAV aerial images. Compared with the traditional methods and
benchmark models, the algorithm has obvious advantages in target detection accuracy and detection
speed.
Songyun Zhang proposed a fast target detection method for UAV imagery based on MobileNet-
YOLO V4 model [9]. The authors used MobileNet as the basic network structure, which is a
lightweight convolutional neural network with fewer parameters and computational complexity,
suitable for target detection on devices with limited resources. Combining MobileNet with YOLO V4,
which is an advanced model in the field of target detection, a series of technological improvements,
such as the CIOU loss function, the SAM module, and PANet, are used to improve the accuracy and

250
 Proceedings of the 2023 International Conference on Machine Learning and Automation
DOI: 10.54254/2755-2721/32/20230219

robustness of target detection. The authors used a series of optimisation measures to further improve
the speed of target detection. For example, the number of parameters of the model is reduced and the
network structure is optimized by network pruning and quantization techniques, which enables the
model to perform target detection in UAV imagery quickly and efficiently. In addition, this paper also
carries out data enhancement and preprocessing for the characteristics of UAV images, which
increases the diversity of samples and improves the generalization ability of the model. Experiments
are conducted on UAV image datasets containing various types of targets and different complex
scenes, and the results show that the method achieves significant performance improvement in the
UAV image target detection task. Compared with traditional methods and other target detection
models, the method based on the MobileNet-YOLO V4 model has obvious advantages in terms of
speed and accuracy.
Xianghong Cheng et al. improved the YOLO V5 algorithm to detect small targets efficiently and
accurately in UAV aerial images [10]. The research team adopted the YOLO V5 algorithm as the basic
framework. Aiming at the low-resolution characteristics of small targets, the team introduced a higher-
level feature pyramid network to enhance the representation of small targets. To suppress the
background interference, this paper adds an attention mechanism, which enables the algorithm to focus
more on the important features of small targets, thus improving the accuracy of detection. To verify
the effectiveness of the improved YOLO V5 algorithm, the researchers constructed a UAV aerial
image dataset containing many small target samples and conducted a series of experiments. The
experimental results show that the improved YOLO V5 algorithm significantly improves UAV aerial
images' performance. Compared with the traditional method and the benchmark model, the algorithm
has obvious advantages in small target detection accuracy and detection speed.

3. Conclusion
In recent years, with the rapid development of UAV technology, UAV image target recognition plays
an increasingly important role in military, civil and industrial fields. Among them, the target
recognition technology based on YOLO (You Only Look Once) algorithm has attracted much
attention. In this paper, we conduct in-depth research on the improvement of the YOLO algorithm for
UAV image target recognition in the past three years, in order to understand the performance of the
algorithm in solving the problems of rotational angle, small target pixels, etc., and to explore the more
detailed enhancement achieved by different versions of YOLO on these improvements.
Our research identified eight important studies that have done a great deal of exploratory work on
the particular challenges of UAV imagery. First, researchers have proposed a series of solutions to the
problem of target rotation, which is prevalent in UAV imagery. Some of these methods are based on
YOLO and introduce a rotation invariance module, which enables the algorithm to better handle
targets with inconsistent rotation angles. These improvements effectively improve the accuracy of
target recognition and enhance the application of UAVs in dynamic environments.
Second, another group of researchers proposed a series of innovative solutions for the problem of
too small target pixels in UAV images. These methods mainly focus on the feature extraction part of
the YOLO algorithm, which effectively enhances the perception of small targets by introducing the
attention mechanism and image pyramid structure. The results show that these improved algorithms
have outstanding performance in recognizing small targets, which greatly improves the detection rate
of UAVs on small targets and provides strong support for dealing with complex and changing practical
application scenarios.
It is worth noting that although all of these studies improved on the YOLO algorithm, they did not
retain the advantages of its inherent light weight and fast detection. This advantage is especially
important for UAV image target recognition in today's demands for efficient computation. These
researches improve the performance and meet the demand for real-time and practicality in practical
applications, enabling UAV technology to work even better in target searching, monitoring and
tracking.

251
 Proceedings of the 2023 International Conference on Machine Learning and Automation
DOI: 10.54254/2755-2721/32/20230219

In addition, some of the research projects have constructed their own datasets to better validate the
performance of the algorithms. By using targeted datasets, these studies can more fully demonstrate
the superiority of their improved algorithms and optimize them for specific scenarios. This trend in
dataset construction has provided UAV image target recognition research with more reliable
evaluation criteria, allowing algorithms to be trained to produce better results and be better adapted to
specific mission requirements.
In summary, research on improving the YOLO algorithm for UAV image target recognition has
made great progress in the past three years. By improving the algorithm for problems such as rotation
angle, too small target pixels, and constructing a customized dataset while retaining the algorithm's
advantages, researchers have made positive contributions to the development of UAV technology.
However, it is also important to realize that the challenges faced by target recognition in the real world
are complex and diverse, and continuous efforts are still needed to further improve the robustness and
accuracy of the algorithms in the future to promote the application of UAV technology in a wider
range of fields.

References
[1] Li Z, Liu X, Zhao Y, Liu B, Huang Z and Hong R 2021 Journal of Visual Communication and
Image Representation 77 103058
[2] Jiang P, Ergu D, Liu F, Cai Y and Ma B 2022 Procedia Computer Science 199 1066–73
[3] Silva L A, Leithardt V R Q, Batista V F L, Villarrubia González G and De Paz Santana J F 2023
IEEE Access 11 62918–31
[4] An J, Putro M D, Priadana A and Jo K-H 2023 2023 IEEE International Conference on
Industrial Technology (ICIT) 2023 IEEE International Conference on Industrial Technology
(ICIT) (Orlando, FL, USA: IEEE) pp 1–6
[5] Li Z, Pang C, Dong C and Zeng X 2023 IEEE Access 11 61546–59
[6] Sahin O and Ozer S 2021 2021 44th International Conference on Telecommunications and
Signal Processing (TSP) 2021 44th International Conference on Telecommunications and
Signal Processing (TSP) (Brno, Czech Republic: IEEE) pp 361–5
[7] Kumar S and Kumar C 2023 2023 International Conference for Advancement in Technology
(ICONAT) 2023 International Conference for Advancement in Technology (ICONAT) (Goa,
India: IEEE) pp 1–5
[8] Chen W, Jia X, Zhu Zh et al. Computer Engineering and Applications 1-11[2023-07-
27].http://kns.cnki.net/kcms/detail/11.2127.TP.20230705.2129.004.html
[9] Zhang Song Yun. Jiangxi Science 2023 41(02) 339-342+355.DOI:10.13990/j.issn1001-
3679.2023.02.020.
[10] Cheng X, Cao Y, Hu Y et al. Flight Control and Detection 2023 6(01) 80-85.

252