Acta Polytechnica https://doi.org/10.14311/AP.2024.64.0059 Acta Polytechnica 64(1):59–67, 2024 © 2024 The Author(s). Licensed under a CC-BY 4.0 licence Published by the Czech Technical University in Prague DIGITAL MODEL OF AUTOMATED UNMANNED AERIAL VEHICLE WITH EDGE COMPUTING APPLICATION FOR RAILWAY RECONNAISSANCE Aigul Tulembayevaa, Shayakhmet Ashirovb, Evgenij Makarovc, Zhexen Seitbattalovd,∗ a Ministry of Defense of the Republic of Kazakhstan, 14 Dostyk Ave., 010000 Astana, Kazakhstan b General Staff of the Armed Forces of the Republic of Kazakhstan, Department of Geoinformation Support, 14 Dostyk Ave., 010000 Astana, Kazakhstan c National Defence University named after the First President of the Republic of Kazakhstan – Elbasi, Department of Scientific Information and Innovation, Esilsky District, 72 Turan Ave., 010000 Astana, Kazakhstan d L. N. Gumilyov Eurasian National University, Department of Computer and Software Engineering, 2 Satbaev St., 010000 Astana, Kazakhstan ∗ corresponding author: sbtl.jeks@gmail.com Abstract. In the last decade, unmanned aerial vehicles have been widely used in various civil and military field operations. Most modern drones are manually controlled by an operator and require decision making, which requires experience gained through training, which is time consuming and expensive, and even with experienced operators, the human factor can never be excluded. In this paper, an algorithm for an automatic flight mode of unmanned aerial vehicles without an operator’s participation has been developed. The proposed algorithm has been considered in the military reconnaissance of railways using edge computing and computer vision approaches to recognise railroads and moving trains for simulated and real cases as well as send report data to the web platform. As a result, the head of the operation received information about the number of carriages and Global Positioning System coordinates of the recognised train to make the necessary decisions. Keywords: Drone, military intelligence, image processing, tracking of the railroad, robotics simulator. 1. Introduction Modern unmanned aerial vehicles (UAVs) are widely used in the world for many applications, from deliv- ering packages to performing combat missions on a battlefield. The most common civil task for UAVs is reconnaissance and searching for missing people over a large area [1]. Reconnaissance with the application of UAVs often uses Full Motion Video to embed spatial information in a video file and ensure the compati- bility with standard geographic information system environments [2]. In military operations, drones are still applied for reconnaissance of enemy positions, weapon stores and military equipment as well as to provide tactical support and humanitarian aid [3]. Modern conflicts have shown that the use of transport infrastructure by enemies, particularly railways, is critical to the provision of combat units participating on the front lines. Therefore, it is quite important to conduct a reconnaissance of the railway [4] to detect moving trains with weapons, heavy equipment, and enemy forces. However, among the frequently encountered techni- cal problems in using UAVs are poor navigation and a high vulnerability to jamming and spoofing [5]. In addition, the travel routes optimisation [6] and minimi- sation of the data refreshment rate [7] are quite critical issues in UAV reconnaissance. The former could be solved by applying the ant colony algorithm [8] and the latter by adding a few drones with a genetic algorithm to observe a point of interest [9, 10]. Nevertheless, the human factor is a key parameter when performing a reconnaissance operation using UAVs [11]. To minimise the impact of the human factor, we proposed to automate the process of UAV flight and detection of reconnaissance targets. The application of artificial intelligence (AI) to solve the automation problem is widespread around the world. For instance, the AI-based controller for the quadcopter Parrot Be- bop 2 showed a high performance and allowed image processing to detect or track targets in real-time [12]. The most used UAV object detection method is YOLO. For example, Sungtae Moon applied the YOLOv5 model, adding the filter pruning method, and as a result, he achieved 40 % parameter reduction as com- pared to the base model [1]. This study aimed to develop an automated flight algorithm for the UAV and detection method of rail- road and moving trains for military reconnaissance, and then send the Global Positioning System (GPS) coordinates of detected targets to the head of the operation for decision making. 59 https://doi.org/10.14311/AP.2024.64.0059 https://creativecommons.org/licenses/by/4.0/ https://www.cvut.cz/en A. Tulembayeva, Sh. Ashirov, E. Makarov, Z. Seitbattalov Acta Polytechnica Properties Mavic 2 Pro Mavic Air Global Hawk ScanEagle Lancaster Hawkeye Phantom Vision 2 Price 1 499 USD 799 USD 93 000 000 USD 100 000 USD 10 000 USD 1 300 USD Flight dura- tion 31 minutes 20 minutes 28 hours 24 hours 45 minutes 25 minutes Operating altitude 8 000 meters 4 000 meters 18 288 meters 5 943.6 meters 121.92 meters Less than 121 meters Speed 72.42 kmph 67.59 kmph 574.53 kmph 91.73 to 111.01 kmph 40.23 kmph 40.23 kmph Weight 0.9 kg 0.43 kg 14 628.3 kg 22 kg 1.36 kg 1.27 kg Table 1. Specifications of reconnaissance drones. 1.1. Related works UAVs often use computer vision methods to detect and inspect the railway infrastructure. For example, Milan Banic applied a Gaussian filter to reduce im- age noise from the background and used a double threshold to determine the boundaries of the rails [13]. Nevertheless, considering the results of the processed images, it can be seen that not the entire length of the railway was recognised based on the proposed algorithm. There were unrecognised areas of railway infrastructure closer to the horizon and to the ob- server. Convolutional Neural Network (CNN) based on U-Net network [14] and Deep CNN methods are also commonly used to track the railway and identify vegetation as obstacles along the railway path [15]. In addition, one of the alternatives for railway recognition is sleeper detection based on the Region-based CNN approach [16]. The method based on the RSNet net- work has shown the best results in railway recognition, reaching 0.973 dice coefficient and 0.94 jaccard [17]. Despite that, all considered methods above require powerful computing resources and discrete graphics cards for complex calculation and object recognition. The next step is to recognise a train. For instance, to solve the problem of detecting traffic jams and recognising crowds of people to assist traffic police, drones with the YOLOv5 model were used. However, the authors noted that the process of recognising small targets was complicated [18], and it should also be noted that the approach consumes significant comput- ing resources and energy of the device. In order to overcome these disadvantages, some researches used edge computing [19, 20]. The experimental results of the study using edge computing significantly opti- mised the wireless network bandwidth and improved scalability without compromising the accuracy of the results and latency [21, 22]. 2. Materials and methods Webots R2023b has been chosen as the 3-dimensional robot simulator [23, 24] to develop an automated flight algorithm for UAV, as it has a wide scope of appli- cations, allows programming the controller, contains a large library of robots and tools for their design, maintains a large number of widely used programming languages, and it is a cross-platform software. As the programming language for the development of the railway and moving trains recognition algorithm, Python has been chosen, as it has advanced and free libraries in the field of computer vision. Thingspeak has been used as a web-based reporting platform since it supports graphical visualisation. Furthermore, we have proposed a general scheme of the system with a description of the reconnaissance process and object recognition of a railway. 2.1. Robot simulator and flight uav algorithm From the existing robot base of Webots, DJI MAVIC 2 PRO has been chosen as the UAV for military in- telligence gathering. The DJI Mavic 2 PRO has the most optimal weight, wind tolerance, and flight dura- tion among the reconnaissance drones. The general specification of reconnaissance drones is presented in Table 1. The drone and designed railway tracks are shown in Figure 1. Figure 1. Mavic 2 pro and railway in Webots. The DJI MAVIC 2 PRO drone comes with a built- in GPS module, a camera, a gyroscope, and lidars. The programming language C was used to develop the drone controller. During the operation of the UAV, the start of the motor, GPS module, camera, 60 vol. 64 no. 1/2024 Digital model of automated unmanned aerial vehicle gyroscope, and lidars were initialised. The built-in object detection method was used to recognise the railway and the train in the simulated environment. The object recognition process is discussed in the next subsection. After detecting the railway, the drone ran an algorithm for following the recognised railroad track with the transmission of GPS coordinates in the console. That process is presented in Figure 2. Figure 2. Railway tracking using UAV and GPS coordinates transmitting. The sections of the detected railway are marked with red rectangles, and there is a console with GPS coordinates at the bottom of the program window. In the case of detecting a train, the drone provided information about GPS coordinates to the head of the operation and marked the train with a green rectangle, as illustrated in Figure 3. Figure 3. The case of detecting a train in Webots. 2.2. Object detection To recognise objects, the OpenCV, Numpy, Matplotlib and Skimage libraries of the Python programming language have been used. Figure 4 demonstrates a flowchart of the object detection process. The result of applying “Equalise Hist” is shown in Figure 5. At the first stage, the colour image was converted into a grey scale image via “COLOR BGR2GRAY”. Figure 4. The flowchart of the object detection process. Figure 5. Equalise Hist image. Then “equalise hist” has been applied to improve the contrast in the image to stretch out the intensity range. The next step was to use Gaussian blur to 61 A. Tulembayeva, Sh. Ashirov, E. Makarov, Z. Seitbattalov Acta Polytechnica Figure 6. The concept scheme of military reconnaissance by UAV. reduce image noise and detail. Otsu’s thresholding has been applied to the obtained result to separate an image into two classes, foreground and background. Further, Canny Edge has been used to locate sharp intensity changes and to find object boundaries in the image. Hough transform has been applied to the final result for detecting simple geometric shapes and lines in images. And, finally, the Hough lines have been combined with the original image. 2.3. General scheme Figure 6 shows a concept scheme of military recon- naissance by UAV of a railway track to represent the process for a real model. The DJ MAVIC 2 PRO transmits GPS coordinates and the camera captures an image of the train and railways, which are then processed by Raspberry Pi 4 Model B (RPi). Computer vision libraries and frame- works have been deployed on the RPi [25, 26]. The results are transmitted to the head of the reconnais- sance operation or a cloud server. 3. Experimental results Returning to the process of object recognition, it is necessary to consider the results of image processing of each stage according to the flowchart shown in Figure 4. The obtained result of converting the image into a grey scale is demonstrated in Figure 7. To provide a clearer explanation, based on the im- age provided below, it is evident that the pixels are clustered around the centre of the available range of intensities. The purpose of Histogram Equalization is to expand this range. A comparison of the Global Thresholding, Otsu’s Thresholding and Otsu’s Thresholding after Gaussian Blur are presented in Figure 8. In the case of global thresholding, we have selected a threshold value arbitrarily. However, the Otsu’s method eliminates the need for manual selection by automatically determining the threshold value. Gaus- sian Blur allowed to get rid of the noise created by the desert sand. The Canny edge detector is a commonly employed technique in computer vision for identifying areas of significant intensity variation and detecting the Figure 7. Converting the original image into COLOR BGR2GRAY. Figure 8. A comparison of the Global Thresholding, Otsu’s Thresholding and Otsu’s Thresholding after Gaussian Blur. boundaries of objects within an image. With this technique, the boundaries of the railway section were determined. The applied canny edge detector to the Otsu’s Thresholding after Gaussian Blur and the fi- nal result of railway detection after Hough Lines is presented in Figure 9. The Hough Line Transform is a transform used to detect straight lines. The Hough Lines have been 62 vol. 64 no. 1/2024 Digital model of automated unmanned aerial vehicle (a). Canny Edge Detection (b). Hough Transform Figure 9. Canny Edge Detector and Hough Trans- form over the original image layer. pasted into the original image. Thereby, the UAV is able to recognise the railroad and trains. The results can be seen in Figure 10. Figure 10. UAV detects the railway and train. It is also important to consider the railway and train recognition algorithm based on a real UAV footage taken from open sources [27]. The recognition algo- rithm for the real case is slightly different from the simulated case due to the huge amount of vegetation in the background, which is shown in Figure 11. The original image is converted to an HSV image Figure 11. The recognition process of the railway for a real case. and forms a mask to highlight the pink-magenta hue to recognise the railway. Next, we followed the standard steps from the previous algorithm, using the Canny Edge Detector and Regions of Interest to display the area of the image with the railroad, and applied the Hough Line Transform to detect straight lines. YOLOv4-tiny has been used for train detection since it supports mobile devices like Raspberry Pi and consumes less computing power than other models. Table 2 shows the benchmarking results for mobile devices, such as Raspberry Pi 4, Jetson Nano, and Rock 5 [28, 29], in frames per second. The result of train recognition is shown in Figure 12. Figure 12. The recognition process of the train for a real case. The mean average precision (mAP) for train recog- nition is 96.67 %. The graph of loss function versus iterations during training is shown in Figure 13. It should be noted that as the distance to the car- riage increases, the mAP decreases. Table 3 shows the mAP for the various distances between the observer and the target object. Thanks to edge computing, the operator and head of the operation received information through the web 63 A. Tulembayeva, Sh. Ashirov, E. Makarov, Z. Seitbattalov Acta Polytechnica Model Size Objects mAP [%] Jetson Nano, FPS RPi 4, FPS Rock 5, FPS NanoDet 320×320 80 20.6 26.2 13.0 36.0 NanoDet Plus 416×416 80 30.4 18.5 5.0 24.9 YoloFastestV2 352×352 80 24.1 38.4 18.8 65.4 YoloV2 416×416 20 19.2 10.1 3.0 20.0 YoloV3 352×352 tiny 20 16.6 17.7 4.4 15.0 YoloV4 416×416 tiny 80 21.7 16.1 3.4 22.4 YoloV4 608×608 full 80 45.3 1.3 0.2 1.5 YoloV5 640×640 small 80 22.5 5.0 1.6 12.5 YoloV6 640×640 nano 80 35.0 10.5 2.7 20.8 YoloV7 640×640 tiny 80 38.7 8.5 2.1 17.9 YoloV8 640×640 nano 80 37.3 14.5 3.1 16.3 YoloV8 640×640 small 80 44.9 4.5 1.47 9.2 Yolo× 416×416 nano 80 25.8 22.6 7.0 28.5 Yolo× 416×416 tiny 80 32.8 11.35 2.8 18.1 Yolo× 640×640 small 80 40.5 3.65 0.9 7.5 Table 2. Benchmark comparison for various mobile devices and computer vision models. Figure 13. The loss function versus iterations during training. 64 vol. 64 no. 1/2024 Digital model of automated unmanned aerial vehicle Figure 14. The report on the web platform ThingSpeak. Figure 15. The network bandwidth in the ‘MonitoriX’ web application’. Distance, meters mAP, per cent 3–7 51.27 7–15 98.19 15–75 90.33 over 75 less than 63.41 Table 3. Dependence of the mean average precision on the distance to the object. platform ThingSpeak about the latitude and longitude of the recognised train, the number of carriages, and the flight time in seconds. The interface of the analytic platform service, channel, widgets, and report are shown in Figure 14. As can be seen from the report results, 6 carriages for transporting people were recognised, the UAV’s flight time was 8.22 seconds, coordinates of the de- tected train: latitude – 46.69335 and longitude – 41.96025. The use of edge computing made it possible to optimize network bandwidth, reaching 13 kilobytes per second or 33 packets per second without errors during data transmission (Figure 15). Thus, the application of edge computing reduces the load on the server, allowing data to be processed on the spot and transfer the final results to the server for reporting and decision making. 4. Discussion Thereby, the application of edge computing has con- tributed to the optimisation of network bandwidth, in contrast to cloud computing, which requires at least 1.5 megabits per second for video streaming [25]. As a result, the proposed algorithm reduced the load on the server and also allowed it to perform other tasks. However, it should be noted that frames per second readings on the Raspberry Pi will be significantly lower than on workstations and personal computers. Comparing the mAPs of our study, it is worth noting that very good results were obtained compared to the accuracy of 35.3 % for the GBS-YOLOv5 [18] and 95.6 % for the all-in-one YOLO [30]. 65 A. Tulembayeva, Sh. Ashirov, E. Makarov, Z. Seitbattalov Acta Polytechnica 5. Conclusions This study has introduced an automated flight al- gorithm for unmanned aerial vehicles (UAVs) and a detection method for identifying railroads and mov- ing trains for simulated and real cases. This research allowed the transmission of the UAV flight time, GPS coordinates, and the number of detected targets to the operation leader. Additionally, the algorithm proposed in this study aims to reduce the time and fi- nancial resources required for operator training, while also minimising human error. Moreover, the scheme of military reconnaissance by UAV has been proposed for the simulated and real cases. As a suggestion for future research, this study could be expanded to include features, such as recognizing types of civilian carriages and armoured trains, and adding an infrared camera for night reconnaissance. The development of a similar system based on an FPGA is being considered to increase the recognition speed. Acknowledgements This research is funded by the Committee of Science of the Ministry of Science and Higher Education of the Republic of Kazakhstan (Project No. BR117020/0221). References [1] S. Moon, J. Jeon, D. Kim, Y. Kim. Swarm reconnaissance drone system for real-time object detection over a large area. IEEE Access 11:23505–23516, 2023. https://doi.org/10.1109/ACCESS.2022.3233841 [2] M. A. Mohamad Sharom, M. F. Ahmad Fauzi, M. Azmi, A. Sipit. Development of video data post-processing technique: Generating consumer drone full motion video (FMV) data for intelligence, surveillance and reconnaissance (ISR). Defence S and T Technical Bulletin 14(1):70–81, 2021. [3] C. Bender, J. Staggs. Leveling the playing field: Equipping Ukrainian freedom fighters with low-cost drone detection capabilities. In 15th International Conference on Cyber Conflict, CyCon 2023, vol. 2023-May, pp. 287–312. 2023. https: //doi.org/10.23919/CyCon58705.2023.10181421 [4] A. O. Lebedev, V. V. Vasil’ev. UAV control algorithm in automatic mode using computer vision. Optoelectronics Instrumentation and Data Processing 57(4):406–411, 2021. https://doi.org/10.3103/s8756699021040075 [5] P. Kozak, M. Vrsecka. The use of drones in military conflict. In 9th International Conference on Military Technologies, ICMT 2023, pp. 1–6. 2023. https://doi.org/10.1109/ICMT58149.2023.10171263 [6] N. Ahn, S. Kim. Optimal and heuristic algorithms for the multi-objective vehicle routing problem with drones for military surveillance operations. Journal of Industrial and Management Optimization 18(3):1651– 1663, 2022. https://doi.org/10.3934/jimo.2021037 [7] D. Kim, L. Xue, D. Li, et al. On theoretical trajectory planning of multiple drones to minimize latency in search-and-reconnaissance operations. IEEE Transactions on Mobile Computing 16(11):3156–3166, 2017. https://doi.org/10.1109/TMC.2017.2689013 [8] L. Ji, C. Zhang, Z. Li, M. Liu. Path planning for drones reconnaissance based on ant colony algorithms. IOP Conference Series: Earth and Environmental Science 330(5):052047, 2019. https://doi.org/10.1088/1755-1315/330/5/052047 [9] J. Tian, L. Shen, Y. Zheng. Genetic algorithm based approach for multi-UAV cooperative reconnaissance mission planning problem. In ISMIS 2006: Foundations of Intelligent Systems, vol. 4203 LNAI, pp. 101–110. Bari, 2006. https://doi.org/10.1007/11875604_13 [10] K. Moldamurat, A. S. Utegen, S. S. Brimzhanova, et al. Development of a software simulator for small satellite swarm control. In 2021 16th International Conference on Electronics Computer and Computation (ICECCO), vol. 2021, pp. 1–5. 2021. https: //doi.org/10.1109/ICECCO53203.2021.9663828 [11] M. Lehto, B. Hutchinson. Mini-drones swarms and their potential in conflict situations. In 15th International Conference on Cyber Warfare and Security, ICCWS 2020, pp. 326–334. 2020. https://doi.org/10.34190/ICCWS.20.084 [12] C. Paucar, L. Morales, K. Pinto, et al. Use of drones for surveillance and reconnaissance of military areas. In MICRADS 2018: Developments and Advances in Defense and Security, vol. 94, pp. 119–132. 2018. https://doi.org/10.1007/978-3-319-78605-6_10 [13] M. Banic, A. Miltenovic, M. Pavlovic, I. Ciric. Intelligent machine vision based railway infrastructure inspection and monitoring using UAV. Facta Universitatis-Series Mechanical Engineering 17(3):357–364, 2019. https://doi.org/10.22190/fume190507041b [14] A. Mammeri, A. Jabbar Siddiqui, Y. Zhao. UAV-assisted railway track segmentation based on convolutional neural networks. In 93rd IEEE Vehicular Technology Conference, VTC 2021-Spring, pp. 1–7. 2021. https://doi.org/10.1109/VTC2021-Spring51267. 2021.9448887 [15] M. A. Rahman, A. Mammeri. Vegetation detection in UAV imagery for railway monitoring. In Proceedings of the 7th International Conference on Vehicle Technology and Intelligent Transport Systems VEHITS - Volume 1, pp. 457–464. 2021. https://doi.org/10.5220/0010439904570464 [16] A. K. Singh, A. K. Dwivedi, M. Sumanth, D. Singh. An efficient approach for instance segmentation of railway track sleepers in low altitude UAV images using mask R-CNN. In IGARSS 2022 – 2022 IEEE International Geoscience and Remote Sensing Symposium, pp. 4895–4898. 2022. https: //doi.org/10.1109/IGARSS46834.2022.9883474 [17] R. S. Rampriya, Sabarinathan, R. Suganya. RSNet: Rail semantic segmentation network for extracting aerial railroad images. Journal of Intelligent and Fuzzy Systems 41(2):4051–4068, 2021. https://doi.org/10.3233/jifs-210349 66 https://doi.org/10.1109/ACCESS.2022.3233841 https://doi.org/10.23919/CyCon58705.2023.10181421 https://doi.org/10.23919/CyCon58705.2023.10181421 https://doi.org/10.3103/s8756699021040075 https://doi.org/10.1109/ICMT58149.2023.10171263 https://doi.org/10.3934/jimo.2021037 https://doi.org/10.1109/TMC.2017.2689013 https://doi.org/10.1088/1755-1315/330/5/052047 https://doi.org/10.1007/11875604_13 https://doi.org/10.1109/ICECCO53203.2021.9663828 https://doi.org/10.1109/ICECCO53203.2021.9663828 https://doi.org/10.34190/ICCWS.20.084 https://doi.org/10.1007/978-3-319-78605-6_10 https://doi.org/10.22190/fume190507041b https://doi.org/10.1109/VTC2021-Spring51267.2021.9448887 https://doi.org/10.1109/VTC2021-Spring51267.2021.9448887 https://doi.org/10.5220/0010439904570464 https://doi.org/10.1109/IGARSS46834.2022.9883474 https://doi.org/10.1109/IGARSS46834.2022.9883474 https://doi.org/10.3233/jifs-210349 vol. 64 no. 1/2024 Digital model of automated unmanned aerial vehicle [18] H. Liu, X. Duan, H. Lou, et al. Improved GBS-YOLOv5 algorithm based on YOLOv5 applied to UAV intelligent traffic. Scientific Reports 13(1):9577, 2023. https://doi.org/10.1038/s41598-023-36781-2 [19] J. Dick, C. Phillips, S. H. Mortazavi, E. de Lara. High speed object tracking using edge computing. In SEC ’17: Proceedings of the Second ACM/IEEE Symposium on Edge Computing, p. 26. 2017. https://doi.org/10.1145/3132211.3132457 [20] T. Kasidakis, G. Polychronis, M. Koutsoubelias, S. Lalis. Reducing the mission time of drone applications through location-aware edge computing. In 2021 IEEE 5th International Conference on Fog and Edge Computing (ICFEC), pp. 45–52. 2021. https://doi.org/10.1109/icfec51620.2021.00014 [21] J. Wang, Z. Feng, Z. Chen, et al. Edge-based live video analytics for drones. IEEE Internet Computing 23(4):27–34, 2019. https://doi.org/10.1109/MIC.2019.2909713 [22] S. K. Atanov, Z. Y. Seitbattalov, Z. S. Moldabayeva. Development an intelligent task offloading system for edge-cloud computing paradigm. In 2021 16th International Conference on Electronics Computer and Computation (ICECCO), pp. 1–6. 2021. https: //doi.org/10.1109/ICECCO53203.2021.9663797 [23] K. Moldamurat, A. Utegen, S. S. Brimzhanova, et al. Development of a software simulator for small satellite swarm control. In 2021 16th International Conference on Electronics Computer and Computation (ICECCO), pp. 1–5. 2021. https: //doi.org/10.1109/ICECCO53203.2021.9663828 [24] Z. Y. Seitbattalov, S. K. Atanov, Z. S. Moldabayeva. An intelligent decision support system for aircraft landing based on the runway surface. In 2021 IEEE International Conference on Smart Information Systems and Technologies (SIST), vol. 2021, pp. 1–5. 2021. https://doi.org/10.1109/SIST50301.2021.9466000 [25] Z. R. Burnayev, D. O. Toibazarov, S. K. Atanov, et al. Weapons detection system based on edge computing and computer vision. International Journal of Advanced Computer Science and Applications 14(5):812–820, 2023. https://doi.org/10.14569/IJACSA.2023.0140586 [26] C. Christakis, K. Ioannidi, S. Sautbekov, et al. The radiation problem from a vertical short dipole antenna above flat and lossy ground: Novel formulation in the spectral domain with closed – form analytical solution in the high frequency regime. Elektronika Ir Elektrotechnika 20(9):35–38, 2014. https://doi.org/10.5755/j01.eee.20.9.8710 [27] Active FPV. DJI fpv max speed 140 km/h vs train - uncut flight[m-mode] || dji action2 footage[2.7k] fpv Thailand. [2023-08-29]. https://www.youtube.com/watch?v=fIzfQJhj86Q [28] Q-engineering. YoloV4 ncnn Raspberry Pi 4. [2023-09-16]. https://github.com/Qengineering/ YoloV4-ncnn-Raspberry-Pi-4 [29] A. S. Utegen, K. Moldamurat, A. Mukhanbetkaliyeva, et al. Development and modeling of intelligent control system of cruise missile based on fuzzy logic. In 2021 16th International Conference on Electronics Computer and Computation (ICECCO), vol. 2021, pp. 1–6. 2021. https: //doi.org/10.1109/ICECCO53203.2021.9663808 [30] Y. P. Wu, P. Chen, Y. Qin, et al. Automatic railroad track components inspection using hybrid deep learning framework. IEEE Transactions on Instrumentation and Measurement 72:1–15, 2023. https://doi.org/10.1109/tim.2023.3265636 67 https://doi.org/10.1038/s41598-023-36781-2 https://doi.org/10.1145/3132211.3132457 https://doi.org/10.1109/icfec51620.2021.00014 https://doi.org/10.1109/MIC.2019.2909713 https://doi.org/10.1109/ICECCO53203.2021.9663797 https://doi.org/10.1109/ICECCO53203.2021.9663797 https://doi.org/10.1109/ICECCO53203.2021.9663828 https://doi.org/10.1109/ICECCO53203.2021.9663828 https://doi.org/10.1109/SIST50301.2021.9466000 https://doi.org/10.14569/IJACSA.2023.0140586 https://doi.org/10.5755/j01.eee.20.9.8710 https://www.youtube.com/watch?v=fIzfQJhj86Q https://github.com/Qengineering/YoloV4-ncnn-Raspberry-Pi-4 https://github.com/Qengineering/YoloV4-ncnn-Raspberry-Pi-4 https://doi.org/10.1109/ICECCO53203.2021.9663808 https://doi.org/10.1109/ICECCO53203.2021.9663808 https://doi.org/10.1109/tim.2023.3265636 Acta Polytechnica 64(1):59–67, 2024 1 Introduction 1.1 Related works 2 Materials and methods 2.1 Robot simulator and flight uav algorithm 2.2 Object detection 2.3 General scheme 3 Experimental results 4 Discussion 5 Conclusions Acknowledgements References