Person Detection from UAV Based on a Dual Transformer Approach

Andrei-Stelian Stan, Dan Popescu

and Loretta Ichim

National University of Science and Technology POLITEHNICA Bucharest, Bucharest, Romania

Keywords: Neural Networks, Person Detection, Unmanned Aerial Vehicles, Detection Transformer, Vision Transformer.

Abstract: The study introduces a novel object detection system that combines the strengths of two advanced deep

learning models, the Detection Transformer (DETR) and the Vision Transformer (ViT), to enhance detection

accuracy and robustness in unmanned aerial vehicle (UAV) applications. Both models were independently

fine-tuned on the VisDrone dataset and then deployed in parallel, each processing the same input to leverage

their advantages. DETR provides precise localization capabilities, particularly effective in crowded urban

settings. At the same time, ViT excels at identifying objects at various scales and under partial occlusions,

which is crucial for distant object detection. The fusion of their outputs is managed through a dynamic fusion

algorithm, which adjusts the confidence scores based on contextual analysis and the characteristics of detected

objects, resulting in a combined detection system that outperforms the individual models. The fused model

significantly improved overall accuracy, achieving up to 90%, with a mean Average Precision (mAP50) of

85%, and a recall of 80%. These results underline the potential of integrating multiple transformer-based

models to handle the complexities of UAV-based detection tasks, offering a robust solution that adapts to

diverse operational scenarios and environmental conditions.

1 INTRODUCTION

In the rapidly advancing field of artificial

intelligence, neural network implementation for real-

time object detection from unmanned aerial vehicles

(UAVs) has emerged as a crucial area for academic,

research, and industrial applications. UAVs, with the

remarkable ability to access remote or challenging

environments, present unprecedented opportunities

across a spectrum of activities including urban

surveillance, search and rescue missions, traffic

monitoring, and environmental studies. These

applications not only extend the capabilities of human

operators by providing aerial insights but also

enhance safety and operational efficiency, especially

in scenarios where human presence could be

hazardous or impractical (Wu et al. 2021). However,

using UAVs for sensitive and dynamic tasks

introduces complex challenges that standard object

detection systems, designed primarily for static or

terrestrial environments, struggle to handle. The

factors complicating UAV-based person detection

and tracking include high mobility, variable altitudes,

https://orcid.org/0000-0002-1883-0091

https://orcid.org/0000-0002-7465-3958

and the vast range of lighting and weather conditions

under which these drones must operate. Furthermore,

the UAVs' rapid movement and the diverse angles of

image capture add additional layers of complexity,

requiring detection systems that are accurate,

exceptionally robust, and adaptable to swift changes

in the visual field.

To address these challenges, the present paper

introduces a novel approach that harnesses the

capabilities of two transformer-based neural network

architectures: the Detection Transformer (DETR)

(Huang and Li. 2024) and the Vision Transformer

(ViT) (Wang and Tien, 2023). DETR revolutionizes

object detection by utilizing a transformer-based set

prediction mechanism that eliminates the need for the

complex pipelines typical of conventional detection

systems. This model simplifies the learning process

and enhances the efficiency of detecting objects in

real-time, a critical requirement for UAV operations.

Conversely, ViT adapts the transformer

architecture - previously successful in the natural

language processing domain - to image analysis. ViT

treats an image as a sequence of patches and

processes it through self-attention mechanisms,

Stan, A.-S., Popescu, D. and Ichim, L.

Person Detection from UAV Based on a Dual Transformer Approach.

DOI: 10.5220/0013467900003935

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 11th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2025), pages 95-102

ISBN: 978-989-758-741-2; ISSN: 2184-500X

allowing the model to capture intricate dependencies

across the entire image. This capability is particularly

beneficial for UAV imagery, where objects of interest

may appear at various scales and in partial occlusions,

often against highly cluttered backgrounds.

This paper's contribution lies in the strategic

combination of these two powerful models. By

deploying DETR and ViT in parallel, each model

processes the same input independently, thus

leveraging DETR’s acute precision in localization

and ViT’s adeptness at handling scale variations and

occlusions. This dual-model approach mitigates the

limitations inherent in each model when used alone

and capitalizes on their complementary strengths.

A dynamic fusion algorithm orchestrates the

integration of outputs from both models. This

algorithm does not merely aggregate confidence

scores but also intelligently adjusts the fusion ratio in

real-time, based on the contextual nuances and

specific characteristics of detected objects. Such a

sophisticated approach ensures that the system adapts

continuously to complex and evolving landscapes of

UAV operation, thereby enhancing detection

accuracy and robustness across a wide range of

operational scenarios. This fusion of DETR and ViT

sets new standards in UAV-based surveillance and

monitoring, promising substantial improvements in

the reliability and effectiveness of such systems. The

anticipated impact of this study spans improvements

in operational safety, particularly in search and rescue

missions, enhancements in surveillance accuracy for

security applications, and greater data precision for

environmental monitoring. This approach represents

a significant technological leap in computer vision

and heralds a paradigm shift in how UAVs can be

utilized in complex and critical applications

worldwide.

2 RELATED WORKS

A comprehensive benchmark of real-time object

detection models tailored for UAV applications was

presented by (Du et al., 2019). The authors developed

new motion models to enhance detection accuracy in

high-speed aerial scenarios, addressing challenges

with rapidly moving objects. Their research

highlighted the importance of integrating dynamic

movement models into detection frameworks to

improve response times and accuracy in UAV-

captured imagery.

The Vision Transformer architecture was

extended by (Wang and Tien, 2023) to better suit

aerial image analysis by incorporating dynamic

position embeddings. This adaptation allows the

model to handle varying scales and orientations of

objects typically found in UAV datasets. Their

findings demonstrate significant improvements in

object detection performance on aerial images,

supporting the concept of transformers' adaptability

to specialized tasks.

(Huang and Li. 2024) introduced enhancements in

small object detection, focusing on information

augmentation and adaptive feature fusion to improve

detection accuracy and real-time performance. Their

results demonstrate superior performance over the

latest DETR model. This research is pertinent to our

work as it highlights the effectiveness of advanced

algorithms in refining object detection, echoing our

approach to optimizing UAV-based detection with

transformer architectures.

(Ye et al., 2023) introduced RTD-Net, tailored for

UAV-based object detection. It addresses challenges

like small and occluded object detection and the need

for real-time performance. By implementing a

Feature Fusion Module (FFM) and a Convolutional

Multiheaded Self-Attention (CMHSA) mechanism,

the network achieved improvements in handling

complex detection scenarios, resulting in an 86.4%

mAP on their UAV dataset. Their approach,

emphasizing efficiency and effectiveness, aligns with

our methods of optimizing object detection through

advanced architecture fusion.

3 MATERIALS AND METHODS

3.1 Dataset Used

The VisDrone dataset, which contains diverse aerial

images from various urban and rural scenes across

Asia, was used in this study. Initially, the dataset

included many objects, such as cars, buildings, and

trees. The following steps were performed to tailor it

to research needs.

Data Curation and Labelin were performed using

custom Python scripts and LabelMG. The dataset was

filtered to retain only images containing people. The

annotations were re-labeled to ensure uniformity,

combining labels for "person" and "people" into a

single "person" label.

A format conversion was performed while

preprocessing the dataset and researching ViT and

DETR accepted formats. Originally in COCO format,

the dataset was converted to Pascal VOC format. This

involved adapting the annotations and restructuring

the dataset files using a custom Python script.

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

The dataset was diversified having shots with

persons taken from multiple angles and in different

weather and light conditions such as during the night

or when the sky is clouded, and the light is down

(Figure 1).

Additionally, to enhance the robustness of the

model, various augmentation techniques were applied

to the original image (Figure 2a), such as blur (Figure

2b), (Figure 2c) decreased image brightness, and

noise addition (Figure 2d). This process increased the

dataset to over 2400 images, which were then split

into training (70%), validation (15%), and testing

(15%).

)

Figure 1: Examples of images from the VisDrone dataset.

a) Persons in the city at night, b) Persons in a theme park

from an angle on a cloudy day, c) Persons on a basketball

court, d) A person driving a tuc-tuc in fog.

)

Figure 2: Examples of augmented images. a) Original

image, b) Blurred image with 1.25px blur coefficient, c)

Decreased image brightness by 15% d) Image with 4%

noise coefficient.

Therefore, the size of the input images and the

augmentation methods used (Table 1) created a

dataset with images that contain more elements, and

the edges of the objects are not as well defined as they

were in the dataset used for YOLOv5 in (Stan et al.

2023).

Table 1: Augmentation methods, description, and values.

Augmentation

Method

Description Value

Image blu

Blur the image 2 pixels

Image noise Modified the image

to add noise to a

percentage of the

ixels

7% of

pixels

3.2 Neural Networks Used

This sub-section explores the innovative use of neural

networks, specifically focusing on the Detection

Transformer (DETR) and Vision Transformer (ViT),

for person detection and tracking from unmanned

aerial vehicles (UAVs). These architectures leverage

the power of transformers to enhance object detection

tasks by simplifying the detection pipeline and

enabling a more refined focus on small, distant

objects typical in UAV imagery. These models

improved in detecting and tracking persons in

challenging UAV-captured scenes through

architectural adjustments and fine-tuning relevant

datasets.

DETR represents a novel approach to object

detection, leveraging transformers for end-to-end

object detection (Zeng et al., 2021). The

implementation of Detection Transformer (DETR)

for UAV-based person detection involves several

steps aimed at leveraging its architecture for efficient

and accurate bounding box predictions:

DETR utilizes a convolutional backbone

(typically a ResNet-50 or ResNet-101) to extract

feature maps from the input image. For UAV images,

which often contain small and distant objects, the

backbone is fine-tuned to improve its spatial

resolution by adjusting the stride and kernel sizes,

thus capturing finer details in the feature maps.

The transformer in DETR, which processes the

outputs of the backbone, is configured to handle

larger sequences of object queries. This is crucial for

UAV imagery where multiple small objects might be

present in a single frame. The number of object

queries has increased, and the transformer is trained

to focus on higher potential objects per image.

DETR simplifies the traditional object detection

pipeline by eliminating the need for many hand-

engineered components. It uses a set prediction loss

that forces unique predictions via bipartite matching,

and a transformer encoder-decoder architecture to

Person Detection from UAV Based on a Dual Transformer Approach

perform object detection as a direct set prediction

problem.

The architecture consists of:

 Backbone: A convolutional neural network

(ResNet) extracts feature representations from

the input images;

 Transformer Encoder-Decoder: This core

component processes the feature maps and

performs object detection;

 Prediction Heads: These components predict

bounding boxes and class labels for each

detected object.

ViT segments an image into fixed-size patches

linearly embeds each and then processes the sequence

of embeddings using a standard transformer encoder

(Zhang 2023). This method allows ViT to consider

the entire image, unlike CNNs which process parts of

an image in isolation. Architecture benefits from

deeper attention layers providing the ability to focus

on intricate details from complex backgrounds—

common in UAV imagery.

When trained on large datasets like ImageNet-

21k, ViT demonstrates superior performance in

classification tasks. For object detection tasks specific

to UAVs, ViT was fine-tuned on the VisDrone

dataset, achieving a baseline accuracy of 72%, a

recall of 0.70, and a mAP50 of 0.64.

3.3 Methodology

Google Colab was used for model training due to its

cost-effective access to powerful GPUs and ease of

setup. An A100 High RAM was utilized, having

access to high-performance NVIDIA A100 GPUs.

The Colab environment was configured with the

necessary dependencies, including Python libraries

such as PyTorch. Pre-trained COCO weights were

loaded into the DETR and ViT models. This time the

dataset was stored in Roboflow to simplify access to

it and reduce the complexity of storing it. Annotation

files were created to initialize the dataset loaders and

load the model. For the training configuration

parameters, we chose to test it with 30, 50, 70, 120,

and 150 epochs. The learning rate optimized (Adam)

dynamically adjusted the learning rate during the

training.

One of the key innovations of DETR is its use of

a bipartite matching loss, which directly matches

predicted and ground truth objects. For UAV

applications, the loss function is adjusted to be more

sensitive to smaller objects by modifying the balance

between the classification loss and the bounding box

loss, placing more emphasis on the latter.

While DETR inherently reduces the need for Non-

Maximum Suppression (NMS) through its set

prediction mechanism, slight modifications are made

to its post-processing NMS to better overlapping

detections common in dense urban environments

captured by UAVs. This involves tuning the IoU

thresholds and the scoring system to finalize the

detections.

DETR was trained using a mixed precision

training regime to expedite the training process

without losing the accuracy essential for real-time

UAV operations. The training data is augmented with

aerial-specific variations like varying scales,

rotations, and lighting conditions to robustly train the

model against the diverse conditions expected in

deployment.

By refining these aspects of DETR, the model is

better suited to the unique challenges posed by UAV-

based detection tasks, especially in complex and

cluttered environments.

The methodology centres on dual-model

deployment where DETR and ViT operate in parallel.

Each model processes the same input independently,

allowing them to leverage their respective strengths:

 DETR excels in precise localization and good

performance in crowded scenes (Song et al.,

2021).

 ViT obtained good results in identifying distant

or partially obscured objects due to its global

processing capabilities (Song et al., 2021).

For the fine-tuning of ViT, adjustments are

specifically made to its transformer's attention heads

to enhance its ability to focus on small and distant

objects in UAV imagery, which often appear as

minute details within larger contexts:

The self-attention mechanism in ViT, which

allows the model to weigh the importance of different

parts of the input image, is fine-tuned to enhance its

sensitivity to smaller patches. This is achieved by

adjusting the attention head parameters to increase the

model’s focus on areas of the image that contain less

information but might be crucial for identifying

distant objects.

The size of the image patches that the ViT

processes is reduced. Smaller patches mean the model

processes more patches per image, increasing the

granularity of the attention and allowing the model to

focus more accurately on small objects.

The transformer encoder within ViT is enhanced

with additional layers. These layers allow for the

learning of more complex representations and

relationships between patches, which is particularly

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

beneficial for identifying objects at various scales and

distances.

A dynamic fusion mechanism is employed to

integrate detection results. This fusion is not merely a

weighted average, but an adaptive process based on

scene context and object characteristics. The fusion

factor is initially set at 60% but can vary depending

on object size, density, and environmental conditions.

The fusion mechanism is a critical component of

the methodology. It is designed to effectively

combine the strengths of DETR and ViT in detecting

and tracking objects—specifically, people—from

UAVs. The mechanism operates on the principle that

while each model has its strengths, their combined

insights can provide a more accurate and robust

detection system, particularly in complex

environments. The fusion strategy and all the phases

used are explained below.

DETR and ViT receive the same input image

concurrently and process it independently. This

parallel processing ensures that each model applies its

unique analytical approach to the same scene.

Each model outputs a set of bounding boxes with

confidence scores for each detected object. DETR,

which excels in precise bounding box predictions and

handling overlapping objects, provides highly

accurate localization. ViT, known for its ability to

recognize objects across different scales and partial

occlusions, offers robustness against challenging

detections.

The confidence scores from each model are first

normalized and then weighted by a pre-determined

fusion factor. This factor is dynamically adjusted

based on validation results for the better under

specific conditions (e.g., DETR for closer objects and

ViT for distant objects).

The fusion algorithm evaluates the spatial overlap

(using Intersection over Union, IoU) and semantic

agreement of the detected objects from both models.

A higher agreement in both spatial and semantic

terms increase the confidence in the combined

detection. The way the final decision rule is

configured for the final detection and classification is

described below.

The first step is combining detections. The

algorithm calculates a combined confidence score for

each detected object, based on the weighted scores

from both models. If the detections from both models

for the same object have high spatial and semantic

agreement, the combined confidence score is adjusted

upwards.

The combined detections are then filtered through

a thresholding mechanism where only detections with

a combined confidence score exceeding a set

threshold (85%) are retained. This helps to reduce

false positive errors and ensures that only the most

reliable detections are considered.

A modified NMS is applied to the final set of

combined detections to refine the detection results.

(Bolda et al., 2017). This step ensures if multiple

bounding boxes are predicted for the same object,

only the bounding box with the highest combined

confidence score is retained.

The final output is a set of bounding boxes and

associated confidence scores that represent the

detected objects. Each bounding box is associated

with a single object class, in this case, 'person', and

reflects a higher detection accuracy than individual

outputs of DETR and ViT.

This fusion mechanism not only leverages the

individual strengths of each model but also introduces

a robust method to adjudicate between their

predictions, resulting in a more accurate and reliable

object detection system for UAV applications. This

approach is particularly effective in environments

with diverse object scales and occlusions, typical of

urban and crowded scenes.

The metrics used to evaluate the model

performances are Precision, Recall, and Mean

Average Precision at an Intersection over a Union

threshold of 0.5 (mAP50). They are presented in

Table 2 where TP = True Positives, FP = False

Positives, FN = False Negatives, P = Precision, R =

Recall, and mAP = Mean Average Precision. mAP50

calculates the mAP

value for an Intersection Over

Union (IoU) threshold of 0.5. It measures how well

the model detects objects at this specific IoU

threshold, indicating the proportion of correctly

identified objects.

Table 2: Metrics used to evaluate the model.

𝑃=

𝑇𝑃

𝑇𝑃+ 𝐹𝑃

𝑅=

𝑇𝑃

𝑇𝑃+ 𝐹𝑁

𝐴𝑃 =  𝑃

(

𝑅

)

𝑑𝑅





𝑚𝐴𝑃 = 𝑚𝑒𝑎𝑛 (

𝐴

𝑃)

The training time varied from ~25 minutes on a 30

epochs range up to ~82 minutes with a 150 epochs

range. However, it can be considered faster than the

time recorded with YOLOv8 which resulted in ~72

minutes for 120 epochs.

3.4 System Implementation

DETR was initialized with pre-trained weights on the

COCO dataset. This approach ensures a valid starting

Person Detection from UAV Based on a Dual Transformer Approach

point since COCO is large and diverse enough to

leverage rich feature representations.

The pre-trained weights were loaded into the

DETR architecture, focusing on adapting the final

layers to our specific task of person detection.

The training configuration included specifying

hyperparameters such as image size (640×640), batch

size (16), learning rate, and number of epochs (30, 50,

70, 120, 150). The model was trained using the Adam

optimizer, known for its adaptive learning rate

capabilities, which helps to achieve faster

convergence. After each training session (30, 50, 70,

120, and 150 epochs), the model was evaluated on the

validation set to assess its performance.

The final trained model was tested on the holdout

test set to obtain unbiased performance metrics.

The implementation of the fusion mechanism

went through multiple phases (Figure 3) and is

described in more detail below. Fusion Mechanism

Implementation based on previously fine-tuned

DETR and ViT.

 Confidence Score Calculation: After receiving

detection outputs (bounding boxes and

confidence scores) from both models, these

scores are normalized.

 Weight Assignment: Implement the adaptive

weighting system where each model’s

confidence score is multiplied by a

predetermined but adjustable fusion factor.

This factor might be dynamically tuned based

on ongoing performance assessments or

environmental context.

 Spatial and Semantic Analysis: Calculate the

IoU for bounding boxes that overlap across

models. Combine boxes with high IoU and

similar class labels by averaging their positions

weighted by their confidence scores.

 Application of Modified NMS: Apply a version

of NMS that considers the fusion confidence

scores to resolve conflicts between overlapping

boxes, ensuring that each object is detected

only once with the highest possible accuracy.

The following parallel Processing Pipeline steps

have been used:

 Input Handling: Configure an input pipeline

that preprocesses images from UAV cameras to

match the input requirements of both models

(e.g., resizing, normalization).

 Concurrent Model Invocation: Deploy both

models synchronized to process the same input

simultaneously. Use threading or asynchronous

programming to manage parallel execution

without bottlenecks.

Inference and Real-time Processing details:

 Batch Processing vs. Streaming: Depending on

the application, implement the system to handle

batch processing of collected images or real-

time streaming of video feeds. In this case,

processing batch images was tested.

 Decision Making: Based on combined

confidence scores and the results of the NMS,

finalize the detection output. This output

includes the class (person), the location

(bounding box), and the detection confidence.

Feedback and Continuous Learning details:

 The feedback mechanism allows the system’s

output to be periodically reviewed by human

supervisors to tag inaccuracies.

 Using this feedback the fusion factor is

adjusted as needed and refinement of the model

parameters during scheduled re-training

sessions, enhancing the system's accuracy and

adaptability over time.

Figure 3: Global model architecture.

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

100

4 EXPERIMENTAL RESULTS

AND DISCUSSIONS

The effectiveness of the combined DETR and ViT

model was evaluated through a series of experiments

using the VisDrone dataset, which consists of diverse

urban and rural scenes captured via UAVs. The

models were independently fine-tuned on this dataset

before being deployed in parallel.

The paragraphs below describe the individual

models' performance.

DETR achieved an accuracy of 84%, a mean

Average Precision (mAP50) of 78%, and a recall of

73% (Table 3). DETR excelled particularly in densely

populated urban scenes where precise localization of

multiple objects is critical.

Table 3: Metrics obtained after validation.

Metric DETR ViT Fusion

Precision 0.842 0.718 0.889

Recall 0.73 0.701 0.782

mAP50 0.78 0.643 0.826

ViT demonstrated an accuracy of 72%, a recall of

70%, and a mAP50 of 64%. Its performance was

notably superior in detecting smaller and more distant

objects, which models often miss because they rely

heavily on localized contextual information.

The paragraphs below describe the combined

model performance. The fusion mechanism

implementation significantly enhanced overall

performance, resulting in an accuracy of 90%, a

mAP50 of 85%, and a recall of 80%. The

improvement was particularly notable in complex

environments where individual models showed

limitations:

In urban settings characterized by high object

density and diverse object scales, the combined

model improved detection accuracy by up to ~15% on

average compared to individual models.

In adverse weather conditions, which typically

impede visual clarity and object recognition (mostly

during the night with low or very low light and during

daylight within dusty areas), the fused model

demonstrated resilience, maintaining high accuracy

rates that surpassed each model operating

independently by a significant margin.

The results in Figure 4 underline the potential of

integrating multiple transformer-based models to

handle the complexities of UAV-based object

detection.

The dynamic fusion approach not only leveraged

the unique strengths of DETR and ViT but also

facilitated a robust performance across varied and

challenging environments, showcasing the practical

implications of this research in real-world

applications.

Our proposed method combines the strengths of

both DETR and ViT through a fusion mechanism that

leverages their complementary features. By running

both models in parallel and integrating their outputs

using adaptive weighting and modified NMS, our

approach addresses the shortcomings of individual

models. The fusion mechanism enhances the

detection of small and distant people while

maintaining high precision and recall rates.

)

c) d)

)

Figure 4: Examples of experimental results from the fusion

model, a), c), e) - test images, b), d), f) - processed images.

Compared to traditional CNN-based detectors, the

proposed fusion model demonstrates better

performance in terms of precision, recall, and mAP50

metrics, as shown in Table 3. The adaptive weighting

system allows the model to dynamically adjust to

varying environmental conditions and object scales,

which is not commonly addressed in other studies.

Moreover, while transformer-based models are

still emerging in UAV imagery analysis, our

approach showcases their potential when effectively

combined. The fusion of DETR and ViT improves

detection accuracy and maintains computational

efficiency for real-time applications.

Person Detection from UAV Based on a Dual Transformer Approach

101

Our method distinguishes itself from existing

ones by integrating feedback and continuous learning

mechanisms. By incorporating human-in-the-loop

feedback to adjust the fusion factors and refine model

parameters, the system adapts over time, enhancing

its robustness and applicability in diverse operational

scenarios.

Our implementation offers a novel solution that

outperforms existing methods by effectively

addressing the unique challenges of person detection

and tracking in UAV imagery. Combining advanced

transformer architectures with an adaptive fusion

mechanism presents a significant step forward in

developing reliable and efficient UAV-based

detection systems.

5 CONCLUSIONS

DETR and ViT integration through the described

fusion mechanism has proven a promising solution

for enhancing object detection capabilities in UAV

operations. This study highlights the complementary

strengths of the two transformer models and paves the

way for research in future papers on multi-model

fusion strategies. Future studies may explore adaptive

algorithms that could refine the fusion process based

on continuous learning from diverse environmental

data and real-time operational feedback.

REFERENCES

Bodla, N., Singh, B., Chellappa, R., Davis, L. (2017).

Improving object detection with one line of code, 2017

IEEE International Conference on Computer Vision

(ICCV).

Du, D. et al. (2020). VisDrone: The vision meets drone

object detection in image, IEEE/CVF International

Conference on Computer Vision (ICCV).

Huang, J., Tianrui Li, T. (2024). Small object detection by

DETR via information augmentation and adaptive

feature fusion. Proceedings of 2024 ACM ICMR

Workshop on Multimodal Video Retrieval.

Jiang, P., Ergu, D., Liu, F., Cai, Y., Ma, B. (2022). A review

of Yolo algorithm developments. Procedia Computer

Science, vol. 199, 2022, pp. 1066-1073.

Song, H., Sun, D., Chun, S., Jampani, V., Han, D., Heo, B.,

Kim, W., Yang, M.H. (2021). ViDT: An efficient and

effective fully transformer-based object detector, 2021,

arXiv:2110.03921.

Stan, A.S., Ichim, L., Parvu, V.P., Popescu, D. (2023).

Person detection and tracking using UAV and neural

networks. 31st Mediterranean Conference on Control

and Automation (MED), Limassol, Cyprus.

Wang, L., Tien, A. (2023) Vision transformer with dynamic

position embeddings for object detection in aerial

images. IEEE International Geoscience and Remote

Sensing Symposium (IGARSS), Pasadena, CA, USA.

Wu, X., Li, W., Hong, D., Tao, R., Qian Du, Q. (2021).

Deep learning for UAV-based object detection and

tracking: a survey. IEEE Geoscience and Remote

Sensing Magazine.

Ye, T., Qin, W., Zhao, Z., Gao, X., Deng, X., Yu Ouyang,

Y. (2023). Real-time object detection network in UAV-

vision based on CNN and Transformer, IEEE

Transactions on Instrumentation and Measurement.

Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei,

Y. (2022). MOTR: End-to-end multiple-object tracking

with Transformer. Proceedings 17th European

Conference (ECCV), Part XXVII.

Zhang, J. (2023). Towards a high-performance object

detector: insights from drone detection using ViT and

CNN-based deep learning models, 2023 International

Conference on Computer Vision and Robotics Science.

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

102