Survey on the Integration of Edge Computing and AI in Drones: Opportunities and Challenges

The following content is sourced from JuShi Security, reprinted from Unmanned Challenge

Unmanned Aerial Vehicles (UAVs), commonly known as drones, face many challenges and applications that can be improved through Artificial Intelligence (AI), edge computing, and edge AI.

AI “describes the workflow of machines that would require intelligence if performed by humans.” Recently, robots/drones have been using AI to complete various intelligent tasks. In contrast to traditional cloud computing, edge computing brings computing services (such as storage and processing) closer to the end users at the network edge. Computing is not performed on remote cloud servers but can be done on nearby edge servers or devices (such as drones/unmanned aircraft), or a combination of both. Edge AI can be summarized as the fusion of AI with edge computing.

The key technical challenges for drones include autonomous navigation, power management, security and privacy, formation control, computer vision, and communication.

Autonomous navigation can be further divided into localization and mapping, path planning, and collision avoidance systems. Important applications of drones include delivery systems, precision agriculture, civil infrastructure inspection, search and rescue operations, serving as aerial wireless base stations, and drone light shows.

Traditional cloud AI computing occurs on cloud servers, providing more processing power and storage compared to edge AI. However, edge AI has many advantages over cloud AI. Notably, it offers lower latency, higher reliability, improved security and privacy, lower costs, and reduced energy consumption.

The identified technical challenges for drones have specific edge AI advantages that are particularly important. For instance, autonomous navigation (especially collision avoidance systems) largely relies on the lower latency of edge AI. Considering the flight speed of drones, if obstacles cannot be detected and avoided within fractions of a second, collisions are likely to occur. The improved latency offered by edge AI compared to cloud AI could mean the difference between a collision and no collision. Different applications also rely more heavily on specific technical challenges. For example, precision agriculture (such as crop management) primarily depends on the technical challenges of computer vision.

While the benefits of edge AI for drones and their technical challenges/applications are indisputable, there are still many challenges in implementing edge AI for drones. Key challenges in implementing edge AI for drones include developing distributed training algorithms, security and privacy, resource allocation, and real-time requirements. However, potential solutions can be identified for each implementation challenge. For example, the implementation challenge of “real-time requirements” outlines that drones often need feedback to be “real-time” (e.g., when tracking fast-moving cars).

However, sometimes collecting data, training models, and deciding actions may take too long. One solution is to reduce inference time through “model pruning” and “knowledge distillation.” Future research directions for edge AI applications in drones can also be defined by the challenges faced and lessons learned.

Drones and edge AI are exciting and active research areas.

The key technical challenges of drone systems can be divided into six categories: 1) Autonomous Navigation; 2) Formation Control; 3) Power Management; 4) Security and Privacy; 5) Computer Vision; 6) Communication

Autonomous navigation can refer to vehicle navigation operated remotely by humans but with some simple onboard algorithms to take over and prevent collisions, or it can refer to fully automated vehicle navigation from point A to point B without any human intervention (e.g., drones for package delivery). Depending on the application, vehicles autonomously navigate by adopting localization and mapping, path planning, and/or collision avoidance. For instance, in the case of remotely operated drones using only some simple onboard algorithms, collision avoidance alone is adopted. In fully autonomous cases, localization and mapping, path planning, and collision avoidance are required. Therefore, this article introduces autonomous navigation from the perspectives of “localization and mapping,” “path planning,” and “collision avoidance systems.”

a) Localization and Mapping:

For robots, mapping is the process of constructing a map of a specific area (2D or 3D), while localization is the process of determining the robot’s position and orientation relative to a reference frame. Accurate localization is a daunting task for any robot, and especially difficult for aerial robots like drones due to the three-dimensional nature of the environment. To make this accurate localization task easier, drones typically rely heavily on the Global Positioning System (GPS), which allows GPS position measurements to be fused with onboard Inertial Measurement Unit (IMU) measurements to accurately estimate the drone’s pose (position and orientation). This method works well because GPS data can compensate for the accumulated errors of the IMU (caused by drift in its measurements).

However, in many cases, GPS services are unavailable/unreliable, such as indoors (factories, warehouses, etc.), in emergencies/disaster situations, near tall trees/buildings, or near water bodies. These environments are aptly called GPS-denied environments, and navigating drones in such environments using only IMU measurements to achieve precise localization is very challenging due to the aforementioned accumulated errors of the IMU.

The most common solution to this problem is a vision-based solution that allows accurate localization of UAVs without GPS by combining IMU and visual sensor measurements. When the two measurements (IMU and visual sensor measurements) are fused, improved pose estimation can be achieved. The most well-known of these vision-based solutions are Visual SLAM (Simultaneous Localization and Mapping) and Visual Odometry (VO). SLAM algorithms aim to estimate the robot’s pose while building a representation of the explored area (visual SLAM algorithms rely on visual sensors). VO incrementally predicts the UAV’s pose by analyzing changes caused by motion in a series of images.

b) Path Planning:

Drone path planning is the problem of determining the path from a starting point to a target point for the drone. There are many types of path planning techniques, but they all attempt to find the optimal (or near-optimal) path based on performance metrics (such as shortest time, shortest route, or lowest operational cost).

As one might expect, the integration of path planning and collision avoidance is crucial. They are often used together, with collision avoidance often referred to as “local path planning,” while path planning is referred to as “global path planning.” The idea is that global path planning generates the best route while considering the entire environment, while local path planning performs localized processing when environmental changes are detected and executes collision avoidance accordingly. Once a collision avoidance operation is executed, an attempt is made to return to the global path.

Drone path planning (global path planning) techniques can be categorized into several different types. One example is the research by Aggarwal and Kumar, which categorizes path planning techniques into representative techniques, cooperative techniques, and non-cooperative techniques.

c) Collision Avoidance Systems:

Because drones fly in the air and often at high speeds, they are at high risk of collisions, whether they collide with other objects or other objects collide with them. Collision avoidance systems are crucial for drones to avoid such collisions and ensure safe flight.

Yasin et al. classify collision avoidance systems into two main categories: 1) Perception and 2) Action, where perception is primary. In the perception phase, sensors perceive the environment and detect obstacles, while in the action phase, collision avoidance methods utilize this information to avoid collisions. Perception is subdivided into active sensors (which emit/launch and then receive their own source) and passive sensors (which only read energy released by other sources). Active sensors are further divided into sonar, LiDAR, and radar, while passive sensors are divided into cameras and infrared.

In the action phase, collision avoidance methods are divided into “four main methods”: 1) Geometric; 2) Optimization; 3) Force Field; 4) Perception and Avoidance. The “Geometric” collision avoidance method uses the velocity and position information of the drone and obstacles to ensure that minimum distance violations do not occur, often involving trajectory simulation. The “Optimization” collision avoidance method uses known obstacle parameters to find the best (or near-optimal) route. The “Force Field” collision avoidance method manipulates attractive/repulsive forces, while the “Perception and Avoidance” method makes avoidance decisions based on obstacle detection in real-time.

The greatest contribution of AI to autonomous navigation is computer vision algorithms, which are crucial for three major challenges: 1) Localization and Mapping; 2) Path Planning; 3) Collision Avoidance Systems. Recent advancements in computer vision algorithms (especially deep learning algorithms) and camera technology (particularly RGB cameras) enable drones to effectively detect objects using high-resolution, lightweight, and inexpensive onboard cameras. In terms of localization and mapping, especially in the absence of GPS data, the use of computer vision algorithms in vision-based methods (such as SLAM and VO) relies heavily on these algorithms (discussed in Section III-A1a). Additionally, in path planning and collision avoidance systems, computer vision algorithms are crucial for rapidly detecting approaching objects to avoid collisions.

Compared to traditional AI under the cloud model, using edge AI can bring many benefits to autonomous navigation. While all the general advantages of edge computing (lower latency, higher reliability, improved security and privacy, reduced costs, and reduced energy consumption) and all the additional advantages of edge AI (further improvements in energy consumption/privacy/communication) apply here, the lower latency and higher reliability provided by edge AI are particularly important for autonomous navigation problems. Therefore, this section briefly discusses the advantages and disadvantages of edge AI.

a) Lower Latency:

Using AI with traditional cloud computing may involve considerable transmission delays in sending and receiving data between the drone and the central cloud server. Edge AI allows most processing to occur on the terminal device or nearby edge servers, resulting in much shorter transmission delays compared to remote clouds. Even in cases described in Section II-B, where some applications require edge servers to communicate with the core cloud for data synchronization, transmission delays are still minimal. This is because preprocessing at the edge reduces the overall traffic load (i.e., less data sent from each device/drone).

Collision avoidance systems are a typical example of a real-time latency-sensitive challenge, which particularly relies on low latency, making edge AI potentially essential. The process of detecting and avoiding approaching objects (unless both the drone and the approaching object are moving very slowly) needs to be completed within fractions of a second. If this process takes too long, it could lead to crashes, resulting in significant economic losses or worse (e.g., fatal consequences). For path planning, localization, and mapping, the latency of traditional AI may be sufficient in most static environments, especially in sparsely populated or high-altitude areas. Nevertheless, considering the limitations imposed by high latency on flight speed, the improved latency provided by edge AI remains a significant advantage.

b) Improved Reliability:

Running AI under traditional cloud computing models relies on the reliability of central cloud servers. If the central cloud server is inaccessible/fails, it could have catastrophic consequences (e.g., crashes). The likelihood of this occurring is much lower with edge AI because processing occurs closer to the user (on edge servers/devices), significantly reducing the chances of network disruptions. If an edge server does fail, another server should be able to provide service; even if not, the terminal device can still handle a large number of requests itself and still access some locally stored data.

When using AI under traditional cloud computing models, the inaccessibility/failure of the central cloud server could lead to crashes as the drone’s video source would not be analyzable, and approaching objects colliding with the drone would not be detectable. With traditional cloud AI, path planning, localization, and mapping can cope with slightly longer times without the central cloud server, but if this connection loss lasts too long, it remains inevitable that the drone will hover in the air, unable to complete its tasks or crash.

c) Disadvantages:

Edge AI has two significant disadvantages that affect autonomous navigation: the apparent lack of network devices in certain areas and fewer technicians to implement, repair, or manage network devices in those areas. In other words, there may be fewer edge servers on the network in sparsely populated areas and/or areas with limited financial or technical resources, and in many such areas, there may be fewer technicians capable of implementing, repairing, and managing edge networks. This can affect autonomous navigation as it may rely on edge servers, meaning the positive impacts of edge AI would be diminished, especially in terms of latency and reliability.

Additionally, another disadvantage for autonomous navigation is the need to offload drone tasks between different edge servers. When a drone leaves the service range of a specific edge server, that server may need to migrate the offloaded tasks of the drone to other edge servers. Given the high speed of drone flight, this may limit the performance of the drone’s tasks.

When executing drone tasks, it is often preferable for multiple drones to work collaboratively rather than a single drone completing the task alone (see the end of Section II-A). Driven by the applications benefiting from the collaborative work of multiple drones (such as transporting large payloads or searching for objects/people over large areas), extensive research on formation-related topics has been conducted in recent decades, with formation control being the most active research topic.

Formation control is inspired by self-organization in nature, such as flocks of birds, and involves the coordinated control of multiple robots in a “formation.” Formation is defined as a network of interconnected agents regulated by controllers, where each agent must maintain relationships with adjacent agents. Formation control is roughly divided into two categories: 1) Leader-Follower and 2) Leaderless.

a) Reinforcement Learning:

Reinforcement Learning (RL) is a field of machine learning, and thus also a field of AI. It can be described as learning how to maximize a numerical reward signal, where the learner is not given actions to take but must try these actions and infer which actions yield the highest rewards. RL is particularly useful for the coordinated movement and communication of drones. As a result, many papers, especially recent ones, have explored the application of RL in drone formation control. For example, Knopp et al. propose an RL algorithm using GQ (λ ) for leader-follower formation control scenarios.

b) Deep Reinforcement Learning:

A problem faced by RL is that it can be very difficult for RL algorithms to learn and determine the optimal reward path (i.e., policy) from each state. RL can be combined with deep learning to produce another subfield of machine learning known as Deep RL (DRL). DRL can use neural networks to handle higher-dimensional state/action spaces. For example, Conde et al. use DRL to drive multiple drones to achieve formation, allowing deep neural networks to estimate the quality of specific states.

c) Multi-Agent Systems and Reinforcement Learning:

Multi-Agent Systems (MAS) is an approach to solving complex problems by breaking them down into smaller tasks. Each individual task is assigned to an autonomous entity called an “agent,” with each agent selecting appropriate actions based on multiple inputs to solve the task. Inputs include its operational history, interactions with neighboring agents, and its objectives. RL can be well applied to such MAS. For instance, Liu et al. propose a distributed model-free solution based on RL for solving the leader-follower formation control problem in MAS.

As with autonomous navigation, formation control benefits greatly from the use of edge artificial intelligence instead of traditional cloud computing models. Thus, most of the general advantages of edge computing (lower latency, higher reliability, etc.) and the general additional advantages of edge AI (e.g., further improvements in energy consumption) apply. Similarly, this section will focus on particularly interesting and important benefits, which for formation control issues is cost reduction.

a) Cost Reduction:

When referring to drone formations, we are talking about swarms of drones that generate a significant amount of streaming data that is continuously transmitted from the drones for processing. Edge AI can reduce the amount of data that needs to be sent to servers through onboard preprocessing and can also reduce the distance data is sent by sending it to edge servers (instead of more remote centralized cloud servers). The reduced amount of data is particularly helpful in lowering operational costs, as cloud computing costs can be quite high when processing large amounts of data. That said, the costs of deploying and maintaining edge devices distributed over vast areas can be quite high and should be considered.

A well-known issue with drones is the limited battery life, which leads to limited flight time. In addition to improving the energy efficiency of drones, more energy can be obtained through innovative means (i.e., energy harvesting technologies) to extend battery life. Common energy harvesting technologies include solar energy, wireless charging, and battery swapping.

a) Solar Energy:

Solar energy is cost-effective and environmentally friendly, making it an ideal energy source. Solar cells are particularly useful for drones because they require long flight times at high altitudes. Solar-powered drones typically use fixed batteries as auxiliary power sources, relying less on batteries during the day but becoming very important at night or in adverse weather conditions. That said, surface area, weight limitations, and dependence on light intensity are major limiting factors for commercial solar-powered drone applications.

b) Wireless Charging:

Wireless charging has not yet become a standard feature for drones. Nevertheless, significant research has been conducted in recent years to make wireless charging for drones feasible. Explored technologies include capacitive coupling-based technologies, magnetic resonance-based technologies, and even charging from power lines through wireless power transmission.

Artificial intelligence can address power management issues through planning and optimization algorithms. For example, algorithms can be employed to find the optimal route for a drone to minimize power consumption or to select the best charging/battery swapping station, considering the current battery level and deviations from the current route. For instance, Zhang et al. used DRL to find the fastest path to a charging station.

Additionally, Convolutional Neural Networks (CNNs) can be used to identify charging stations and their landing points, while Recurrent Neural Networks can accurately predict when a drone will run out of battery, and AI can also help improve the efficiency of drones, allowing them to achieve more goals with less power.

As with Sections III-A3 and III-B3, the general advantages of edge computing and the general additional advantages of edge AI also apply here. The advantage of edge AI in reducing energy consumption is particularly important and will be discussed.

a) Reducing Energy Consumption:

As mentioned in Section III-B3, edge AI can reduce the amount of data that needs to be sent to servers through onboard preprocessing and can also reduce the distance data is sent by sending it to edge servers (instead of more remote centralized cloud servers). In addition to lowering costs, this can also reduce energy consumption, as transferring data to remote clouds can consume a lot of energy. By reducing the size of data and the distance of transmission, overall energy consumption can be reduced, which is especially important for drones since batteries are a scarce resource. Furthermore, when using edge AI, distributed federated learning within drone swarms can further reduce the amount of data sent to servers and the associated energy consumption. For example, Zeng et al. proposed a framework for drone swarms based on distributed federated learning, which involves a “leading drone” and several “following drones.” All following drones locally train models based on their collected data, and then each following drone sends this locally trained model to the leading drone. This drone aggregates all models to create a global model, which it then sends to subsequent drones.

Security and privacy are very important issues for all digital systems, especially for drones. Compared to other privacy-infringing devices, drones have unique flexible access technologies (which are attractive to criminals) and are also vulnerable to attacks targeting wireless links, network elements, physical elements, and interfaces between network and physical elements.

a) Security:

While many papers discussing drone security address the same issues, they take different approaches.

Shakhatreh et al. classify the attack vectors against drone systems into attacks on communication links, the drones themselves, ground control stations (GCS), and humans. They also define three major cybersecurity challenges for drone systems: 1) Confidentiality challenges (referring to protecting information from unauthorized access); 2) Availability challenges (ensuring that drone system services and data work as intended and are accessible to authenticated users); 3) Integrity challenges (ensuring the authenticity of information). Ullah et al. briefly explain the classification of attackers (internal/external, malicious/benign, active/passive, and local/distributed) and attacks (e.g., spoofing, denial of service, and GPS spoofing). However, this section analyzes drone attacks from the perspective of sensors and communication links.

The information obtained by sensors can alter drone behavior and has significant security implications. GPS sensors are a common attack vector, as they are typically used to obtain accurate location information. The most common ways to attack drone GPS include “jamming” or “spoofing.” Jamming refers to unauthorized parties broadcasting interference signals to prevent the reception of various signals. Spoofing can involve unauthorized parties recording satellite signals and transmitting them to drones (relay-type spoofing) or generating signals based on real signals using certain programs (generation-type spoofing). One method to address GPS jamming is to adopt alternative navigation methods, such as using vision and inertial navigation systems that employ SLAM or VO (see the localization and mapping section, Section III-A1a). Solutions to GPS spoofing include authenticating GPS signals, such as checking GPS observables that represent signal travel time or detecting sudden changes in signal power/observables that may indicate the onset of spoofing attacks. Other commonly vulnerable sensors include binocular vision sensors that may be deceived (e.g., inducing features using ground-pointing lasers) or MEMS gyroscopes that may be attacked by ultrasound (e.g., causing the drone to lose balance).

In terms of communication links, the interaction between drones and ground stations relies on communication links that are susceptible to attacks. For example, one way drones communicate with ground terminals is through WiFi, which is vulnerable to Deauth attacks, causing disconnections between drones and terminals, and drone control passwords to be compromised. Solutions include having drones query users to automatically return home after losing connection for 10 seconds; using radio signals instead of WiFi signals.

b) Privacy:

Drone privacy violations are easy to perpetrate and difficult to capture. Preventing drones from infringing on personal privacy primarily involves two solutions. One is to register home addresses in a no-fly zone database, but this still does not guarantee that invading drones won’t fly into no-fly zones. The second solution is to use technologies/systems to detect, track, and deploy drones in space.

Additionally, malware can exploit drones to collect personal information. For example, Snoopy malware can be installed on drones to collect personal information and track/analyze individuals using smartphones. Drones need to continuously evolve to counteract this ever-evolving malware.

Artificial intelligence can be used to help prevent drones from being attacked, as well as to prevent drones themselves from infringing on people’s privacy (by flying over/near no-fly zones). Regarding the latter, by 2026, the “global anti-drone market” is expected to reach $ 25.97 billion. An example is the surveillance system from Japan Electric Company, which uses acoustic, thermal, infrared, and/or radio communication sensors/detectors to sense invading drones and provides options for tracking system owners to drop/capture drones. Algorithms for detecting drones particularly employ artificial intelligence, such as the drone detection algorithm based on artificial neural networks (ANN) proposed by Zhang et al.

For the former (preventing drones from being attacked), Challita et al. argue that it is crucial to detect potential attacks by discovering bad/abnormal drone movements. Challita et al. use RNNs as an example, where RNNs can predict drone movements and thus detect abnormal movements of drones. Furthermore, AI protection can utilize models trained on malicious files to protect drones from zero-day attacks.

As with the “Role of Edge AI” sections for other technical challenges, the general advantages of edge computing and the general additional advantages of edge AI apply. Clearly, the security and privacy advantages of edge AI are somewhat related, and thus will be discussed.

a) Improved Security:

Typically, using AI in traditional cloud computing models requires sending all data to a central server. This is considered “highly vulnerable,” as a single attack (e.g., a DDoS attack) can cause significant damage. Using edge AI means distributing data processing across multiple devices/servers. Although it must be acknowledged that this distribution of data processing increases the potential attack surface, the risk is more dispersed, thus reducing the impact of successful attacks (such as DDoS attacks). Additionally, because edge AI supports processing at the edge, less data is sent, meaning less data is intercepted.

In terms of practical shutdown attacks, the distributed and decentralized nature of edge computing means that the vulnerable parts of edge AI systems are easier to shut down than traditional cloud AI systems, which typically require shutting down the entire network.

Integrating blockchain into drone systems can bring numerous security advantages. For instance, blockchain can ensure that each drone has a copy of the blockchain, thereby reducing the vulnerability of the drone network to signal interference. In this case, individual drones use the detailed flight paths of other drones contained in their blockchain copies to determine their own paths. Another example is how blockchain can detect malicious drones that change network information. A malicious drone may initially be part of the drone network and later be hijacked or enter the network later. Blockchain can help prevent such attacks through blockchain consensus algorithms, where any drone can report suspicious activity. If the number of entries contradicting a drone exceeds a certain threshold, that drone is labeled as malicious.

b) Improved Privacy:

Edge AI enables real-time computation. If it is inevitable to capture/record privacy-infringing images, data can be processed locally on the drone or edge server without being sent and stored on a remote centralized cloud server. The significance of this is that such data is less likely to be hacked and does not need to be stored on a remote centralized cloud server.

Federated learning and blockchain can both enhance the security of drone communications. Federated learning can prevent the transmission of any raw data from devices/drones (only local model updates are sent). Blockchain can encrypt data and store it in the blockchain, making it accessible only to those with the correct decryption keys. Additionally, blockchain can protect four main types of data in the drone network (drone identifiers, flight route control, sensor data, and flight schedules) by writing and updating them within blockchain blocks.

c) Disadvantages:

As mentioned in the security section (Section III-D3a), it should be noted that distributing data/data processing across multiple edge nodes has the disadvantage of increasing the potential attack surface. Even with blockchain-based systems, each drone has a copy of the distributed ledger, meaning some sensitive information spreads across all drones in the system. Additionally, it should be noted that in blockchain-based drone swarms, if more than half of the drones in the swarm are hacked, the entire swarm can be controlled (this type of attack is known as a 51% attack).

The purpose of computer vision is to enable computers to understand the environment from visual information (whether from a single image or a series of images). In recent years, there has been growing interest in automatically understanding the visual data collected by drones, with computer vision playing a crucial role in most drone applications (from aerial photography to SAR operations).

From the perspective of computer vision, the core task of such applications is scene parsing. Different applications require different levels of scene parsing, from locating objects to determining the precise boundaries of objects and recognizing objects. Drone computer vision applications include object detection, object recognition, object tracking, collision avoidance, autonomous navigation, and 3-D reconstruction. Such image processing can be performed remotely on servers (edge or central cloud) or on the drone (embedded).

a) Remote Computer Vision Processing:

Drones often lack the processing power to handle images captured by their cameras, so processing needs to occur at different locations. From a latency perspective, the ideal processing location is an edge server, although computer vision can also be processed on more distant centralized servers.

b) Real-Time Embedded Computer Vision Processing:

If the goal is to enable drones to achieve true autonomy and reliability, real-time embedded computer vision processing is preferable to remote computer vision processing, as remote processing requires high bandwidth, minimal latency, and extremely reliable wireless links, which are not always guaranteed.

The most prominent limitation of real-time embedded computer vision processing for drones is onboard computational capability. Van Beeck et al. point out that the computational demands of state-of-the-art drone computer vision algorithms often conflict with hardware resource limitations.

In computer vision technology, the use of artificial intelligence is not mandatory; for example, Petricca et al. can perform rust detection based on the number of pixels containing certain red components. Nevertheless, despite the fact that AI-based technologies may require large datasets to achieve optimal results, computer vision applications extensively use artificial intelligence. The area of artificial intelligence that intersects most with computer vision is deep learning.

a) Deep Learning:

According to the research by Lecun et al., deep learning enables computational models composed of multiple processing layers to learn data representations with multiple levels of abstraction. The first few layers aim to learn the detection of low-level features such as edges, while the later layers aim to combine features into more complete representations. One example of the application of deep learning in drone computer vision is the method proposed by Ye et al., which uses deep learning classifiers to detect and track other drones. Another example is the method proposed by Padhy et al., which employs CNN models to help drones navigate autonomously in indoor corridor environments without GPS.

The role of edge AI in computer vision involves computer vision processing on edge servers and drones. When AI processing can be completed on edge servers or devices (instead of in traditional cloud models), all the general advantages of edge computing and the general additional advantages of edge AI apply. The advantage of low latency is particularly important for computer vision applications, and thus will be emphasized. As with the role of AI sections (Section III-E2), this section will also particularly highlight deep learning, focusing on embedded deep learning, which has not yet been discussed.

a) Lower Latency:

As mentioned in Section III-A.6, drones using traditional cloud AI experience considerable transmission delays when sending and receiving data between themselves and central cloud servers. In contrast, edge AI allows most processing to occur on terminal devices or nearby edge servers, resulting in significantly shorter transmission delays compared to remote clouds. Even in cases described in Section II-B, where some applications require edge servers to communicate with the core cloud for data synchronization, the overall traffic load is smaller due to preprocessing at the edge, so transmission delays are not significant.

Most computer vision applications (from object detection to object recognition to object tracking to collision avoidance to self-navigation) require low latency because if certain latency standards are not met, objects can be lost/collide.

b) Deep Learning:

According to Van Beeck’s statements at the “UAVision2020” workshop focused on real-time image processing on drones, all accepted workshop papers (covering a wide range of different applications) utilized deep learning. Castellano et al. and Zhao et al. are good examples of using deep learning for crowd counting or understanding. Other examples include Stad and Zhang et al. using deep learning for object tracking, and Peralta et al. using deep learning for 3-D reconstruction. Recent examples include Onishi and Ise using CNN methods to build tree recognition and mapping systems using drone RGB images, and Kung et al. proposing a CNN model for automatic detection of building defects (such as cracks).