In this stage of V-SLAM, we systematically prepare input data using system hardware, which includes capturing and preparing images. It involves installing cameras such as RGB-D cameras, depth cameras, or infrared sensors for collecting data and initializing the system (Beghdadi and Mallem, 2022).

The system gathers data, with a particular emphasis on crucial filtering details aimed at effectively eliminating any noise present in the input data (Mane et al., 2016; Grisetti et al., 2007). The refined data are then sent to the next stage for further processing to extract features from the input information (Ai et al., 2021). As a result, progress in SLAM methods has resulted in the creation of numerous datasets accessible to researchers to evaluate V-SLAM algorithms (El Bouazzaoui et al., 2021).

In the second stage of V-SLAM, the system focuses on finding its location, which is an important part of the entire process (Scaradozzi et al., 2018). It involves the execution of various processes that are crucial for successfully determining where the robot is. Feature tracking plays a central role during this phase, with a primary focus on tasks such as feature extraction, matching, re-localization, and pose estimation (Picard et al., 2023). It aims to align and identify the frames that guide the estimation and creation of the initial keyframe for the input data (Ai et al., 2021). A keyframe is a set of video frames that includes a group of observed feature points and the camera’s poses. It plays an important role for the tracking and localization process, helping in eliminating drift errors for camera poses attached to the robot (Sheng et al., 2019; Hsiao et al., 2017). Subsequently, this keyframe is sent for further processing in the next stage, where it will be shaped into a preliminary map, a crucial part for the third stage of the workflow (Aloui et al., 2022; Zhang et al., 2020).

The third stage of the V-SLAM workflow focuses on the crucial task of building the map, an essential element in V-SLAM processes. Various types of maps can be generated using SLAM, including topological maps, volumetric (3D) maps, such as point cloud and occupancy grid maps, and feature-based or landmark maps. The choice of the map type is based on factors such as the sensors employed, application requirements, environmental assumptions, and the type of dataset used in robotic applications (Taheri and Xia, 2021; Fernández-Moral et al., 2013). In robotics, a grid map is a representation of a physical environment, with each cell representing a particular location and storing data comprising obstacles, topography, and occupancy. It functions as a fundamental data structure for several robotics navigation and localization techniques (Grisetti et al., 2007). A feature-based map is a representation which captures the features of the environment, such as landmarks or objects, to facilitate localization and navigation tasks (Li et al., 2022a). A point cloud map is a representation of a physical space or object made from lots of 3D dots, showing how things are arranged in a place. It is created using special cameras or sensors and helps robots and computers understand what is around them (Chu et al., 2018).

After setting up keyframes during the localization stage, the workflow progresses to field modeling. Then, key points and feature lines are identified and detected, which is crucial for generating a map (Schneider et al., 2018). It is a process that builds and updates the map of an unknown environment and is used to continuously track the robot’s location (Chen et al., 2020). It is a two-way process that works together with the localization process, where they depend on each other to achieve SLAM processes. It gathers real-time data about the surroundings, creating both a geometric and a visual model r13 (accessed on 14 November 2023). In addition, the process includes the implementation of bundle adjustments (BAs) to improve the precision of the generated map before it is moved to the final stage (Acosta-Amaya et al., 2023). BA is a tool that simultaneously refines the parameters essential for estimating and reconstructing the location of observed points in available images. It plays a crucial role in feature-based SLAM (Bustos et al., 2019; Eudes et al., 2010).

The final stage in the V-SLAM workflow involves fine-tuning the process and closing loops, resulting in the optimization of the final map. In V-SLAM, the loop closure procedure examines and maintains previously visited places, fixing any errors that might have occurred during the robot’s exploration within an unknown environment. These errors typically result from the estimation processes performed in earlier stages of the SLAM workflow (Tsintotas et al., 2022; Hess et al., 2016). Loop closure and process tuning can be done using different techniques, such as the extended Kalman filter SLAM (EKF-SLAM). EKF-SLAM combines loop closure and landmark observation data to adjust the map in the Kalman filter’s state estimate. This tool helps address uncertainties in the surrounding world (map) and localize the robot within it (Song et al., 2021; Ullah et al., 2020).

The bag-of-words (BoW) approach is another technique used to enable robots to recognize and recall previously visited locations. This is similar to how humans remember places they have been to in the past, even after a long time, due to the activities that took place there. BoW works by taking the visual features of each image and converting them into a histogram of visual words. This histogram is then used to create a fixed-size vector representation of the BoW, which is stored for use in matching and loop-closing processes (Cui et al., 2022; Tsintotas et al., 2022).

Finally, graph optimization is used as a correction tool for loop closure processes. It refines the final map and robot’s trajectory by optimizing the graph based on landmarks. This technique involves a graph-based representation of the SLAM issue, where vertices represent robot poses and map characteristics and edges represent constraints or measurements between the poses. It is commonly used as a correction tool in graph-based SLAM types (Zhang et al., 2017; Chou et al., 2019; Meng et al., 2022).

In conclusion, these comprehensive workflow processes outlined in Sections 2.1, 2.2, 2.3, and 2.4, respectively, play an important role in V-SLAM for robotics as they facilitate the simultaneous creation of maps and real-time location tracking within the operational environment (Li et al., 2022b).

It is a SLAM system designed to map the environment around the sensors while simultaneously determining the precise location and orientation of those sensors within their surroundings. It relies entirely on visual data for estimating sensor motion and reconstructing environmental structures (Taketomi et al., 2017).

It uses monocular, RGB-D, and stereo cameras to scan the environment, helping robots map unfamiliar areas easily. This approach has attracted attention in the literature because it is cost-effective, easy to calibrate, and has low power consumption in monocular cameras while also allowing depth estimation and high accuracy in RGB-D and stereo cameras (Macario Barros et al., 2022; Abbad et al., 2023). The methods used in this part can be listed herein.

VI-SLAM is a technique that combines the capabilities of visual sensors, such as stereo cameras, and inertial measurement sensors (IMUs) to achieve its SLAM objectives and operations (Servières et al., 2021; Leut et al., 2015). This hybrid approach allows a comprehensive modeling of the environment, where robots operate (Zhang et al., 2023). It can be applied to various real-world applications, such as drones and mobile robotics (Taketomi et al., 2017). The integration of IMU data enhances and augments the information available for environment modeling, resulting in improved accuracy and reduced errors within the system’s functioning (Macario Barros et al., 2022; Mur-Artal and Tardós 2017b). The methods and algorithms used in this approach, while implemented in real-life applications, can be listed as shown in the following section.

RGB-D is an innovative approach that integrates RGB-D cameras with depth sensors to estimate and build models of the environment (Ji et al., 2021; Macario Barros et al., 2022). This technique has found applications in various domains, including robotic navigation and perception (Luo et al., 2021). It demonstrates efficient performance, particularly in well-lit indoor environments, providing valuable insights into the spatial landscape (Dai et al., 2021).

The incorporation of RGB-D cameras and depth sensors enables the system to capture both color and depth information simultaneously. This capability is advantageous in indoor applications, addressing the challenge of dense reconstruction in areas with low-textured surfaces (Zhang et al., 2021b). The objective of RGB-D SLAM is to generate a precise 3D reconstruction for the system surroundings, with a focus on the acquisition of geometric data to build a comprehensive 3D model (Chang et al., 2023). The methods used in this section are listed as follows:

RTAB-Map SLAM, which stands for real-time appearance-based mapping, is a visual SLAM technique that works with RGB-D and stereo cameras (Ragot et al., 2019). It is a versatile algorithm that can handle 2D and 3D mapping tasks depending on the sensor and data that are given (Peter et al., 2023; Acosta-Amaya et al., 2023). It integrates RGB-D and stereo data for 3D mapping, enabling the detection of static and dynamic 3D objects in the robot’s environment (Ragot et al., 2019). It is applicable in large outdoor environments where LiDAR rays cannot reflect and manage the field around the robot (Gurel, 2018). Variable lighting and environmental interactions can cause robotic localization and mapping errors. Therefore, RTAB’s robustness and adaptability to changing illumination and scenes enable accurate operation in challenging environments. It can handle large, complex environments and is quickly adaptable to work with multiple cameras or laser rangefinders (Li et al., 2018; Peter et al., 2023). Additionally, the integration of T265 (Intel RealSense Camera) and implementation of ultra-wideband (UWB) (Lin and Yeh, 2022) address robot wheel slippage with drifting error handling, enhancing system efficiency with precise tracking and 3D point cloud generation, as done in Persson et al. (2023); see Table 2.

The RTAB-MAP SLAM method involves a series of steps that enable it to function (Gurel, 2018; Labbé and Michaud, 2019). Initially, the hardware and front-end stage is responsible for tasks such as obtaining data from stereo and RGB-D cameras, generating frames, and integrating sensors. This stage prepares the frames that will be used in the subsequent stage. After the frames have been processed simultaneously with the tracking process, the loop closure is activated to generate the necessary odometry. Subsequently, the keyframes equalization and optimization processes are initiated to improve the quality of the 2D and 3D maps generated for SLAM applications, as shown in Figure 5, part 7.

The TUM RGB-D dataset is a widely used resource in the field of V-SLAM, which helps demonstrate the effectiveness and practicality of V-SLAM techniques. This dataset provides both RGB images and depth maps, with the RGB images saved in a 640 × 480 8-bit format and the depth maps in a 640 × 480 16-bit monochrome (Chu et al., 2018). It offers RGB-D data, making it appropriate for both depth-based and V-SLAM techniques. Its usefulness extends to essential tasks such as mapping and odometry, providing researchers with a considerable volume of data for testing SLAM algorithms across diverse robotic applications (Ji et al., 2021; End et al., 2012). The adaptability of these datasets is remarkable, as they find application in mobile robotics and handheld platforms, demonstrating effectiveness in both indoor and outdoor environments (Martínez-Otzeta et al., 2022; Son et al., 2023).

Some of the recent studies used TUM datasets, such as in Li et al. (2023c). They have leveraged the TUM RGB-D dataset to establish benchmarks customized to their specific research objectives. The study initiated its investigations with RGB-D images and ground truth poses provided by the TUM datasets, utilizing them to construct 3D scenes characterized with real space features. The integrative role assumed by the TUM RGB-D dataset in this context attains profound significance as a fundamental resource within the domain of V-SLAM research. For more details, refer to the TUM RGB-D SLAM dataset.

The EuRoC MAV benchmark dataset is specifically designed for micro aerial vehicles (MAVs) and contributes a valuable resource in the domain of MAV-SLAM research since it includes sensor data such as IMU and visual data such as stereo images. These datasets, published in early 2016, are made accessible for research purposes and offer a diverse usability in indoor and outdoor applications. Consequently, it serves as a relevant choice for evaluating MAV navigation and mapping algorithms, particularly in conjunction with various visual V-SLAM methodologies (Sharafutdinov et al., 2023; Leutenegger, 2022; Burri et al., 2016).

The EuRoC MAV benchmark dataset, of notable benefits to robotics, is particularly valuable for researchers working on visual-inertial localization algorithms like OpenVINS (Geneva et al., 2020; Sumikura et al., 2019) and ORB-SLAM2 (Mur-Artal and Tardós, 2017a). This dataset incorporates synchronized stereo images, IMU measurements, and precise ground truth data, providing comprehensive resources for algorithm development. Its comprehensive data structure makes it highly suitable for thoroughly testing and validating algorithms tailored for MAV purposes (Burri et al., 2016). For more details, refer to the EuRoC MAV dataset.

The KITTI dataset is a widely utilized resource in robotics navigation and SLAM, with a particular emphasis on V-SLAM. Designed for outdoor SLAM applications in urban environments, KITTI integrates data from multiple sensors, including depth cameras, lidar, GPS, and inertial measurement unit (IMU), contributing to the delivery of precise results for robotic applications (Geiger et al., 2013). Its versatility extends to supporting diverse research objectives such as 3D object detection, semantic segmentation, moving object detection, visual odometry, and road-detection algorithms (Wang et al., 2023; Raikwar et al., 2023).

As a valuable asset, researchers routinely rely on the KITTI dataset to evaluate the effectiveness of V-SLAM techniques in real-time tracking scenarios. In addition, it serves as an essential tool for researchers and developers engaged in the domains of self-driving cars and mobile robotics (Geiger et al., 2012; Ortega-Gomez et al., 2023). Furthermore, its adaptability facilitates the evaluation of sensor configurations, thereby contributing to the refinement and assessment of algorithms crucial to these fields Geiger et al. (2013). For more details, refer to the KITTI Vision Benchmark Suite.

The Bonn dataset is purposefully designed for RGB-D SLAM, containing dynamic sequences of objects. It showcases RGB-D data accompanied by a 3D point cloud representing the dynamic environment, which has the same format as TUM RGB-D datasets (Palazzolo et al., 2019). It covers both indoor and outdoor scenarios, extending beyond the boundaries of controlled environments. It proves valuable for developing and evaluating algorithms related to tasks such as robot navigation, object recognition, and scene understanding. Significantly, this dataset is versatile enough to address the complexities of applications used in light-challenging areas (Soares et al., 2021; Ji et al., 2021). In addition, it proves to be an important resource for evaluating V-SLAM techniques characterized by high dynamism and crowds where the robot might face the challenge of object detection and interaction with the surrounding environment (Dai et al., 2021; Yan et al., 2022). For more details, refer to the Bonn RGB-D dynamic dataset.

It is a benchmark dataset which is designed for RGB-D applications, serving as a valuable tool for evaluating RGB-D, visual odometry, and V-SLAM algorithms, particularly in indoor situations (Handa et al., 2014). It includes 3D sensor data and ground truth poses, facilitating the benchmarking of techniques related to mapping, localization, and object detection in the domain of robotic systems. Its pre-rendered sequences, scripts for generating test data, and standardized data formats are beneficial for researchers in evaluating and improving their SLAM algorithms (Chen et al., 2020). A unique aspect of the ICL-NUIM dataset is its inclusion of a three-dimensional model. This feature empowers researchers to explore and devise new scenarios for robotic systems, which operates in unknown environments. Moreover, it promotes improvements in V-SLAM, which makes it possible to generate semantic maps that improve robots’ flexibility and adaptability to integration into that environment easily and flexibly (Zhang et al., 2021a). For more details, refer to the ICL-NUIM dataset.

When choosing among V-SLAM methods, a key consideration is the robustness and accuracy of the method (Zhu et al., 2022). In particular, a robust algorithm can handle sensor noise, obstacles, and changing environments to ensure continuous and reliable operation (Bongard, 2008). Additionally, accuracy is equally important for creating precise maps and localization, allowing the robot to make informed decisions and move through the environment without errors (Kucner et al., 2023; Nakamura et al., 2023). These qualities collectively enhance the algorithm’s reliability in challenging real-world situations, making them crucial factors for successful mobile robotic applications.

In the application of mobile robotics, the selection of the SLAM algorithm is extremely important, focusing on the efficiency of the process happening inside the robot’s computational architecture (Macario Barros et al., 2022). Therefore, the chosen V-SLAM algorithm must be carefully tailored to meet the computational demands imposed by the real-time constraints of the robot. This entails a delicate balancing act as the selected algorithm should be seamlessly integrated with the available processing power and hardware resources, all while satisfying the stringent real-time requirements of the application. The critical consideration for this step is the quality of the sensors, the professors, and/or computers so that they can generate a quick response and accurate localization in a very limited time (Henein et al., 2020).

In robotic applications, it is important for researchers to choose a SLAM algorithm that works well with the robot’s sensors. Integrating suitable hardware improves speed and performance in SLAM systems through accelerators, method optimization, and energy-efficient designs (Eyvazpour et al., 2023). Various V-SLAM algorithms are designed for specific sensor types such as RGB-D, lidar, and stereo cameras. This facilitates seamless integration into the SLAM system, enhancing the functionality of utilizing integrated hardware (Wang et al., 2022). Moreover, the availability of ROS packages and open-source software for sensors and cameras provides increased modality and flexibility during system installation. This, in turn, enhances adaptability and makes integration easy and free of challenges (Sharafutdinov et al., 2023; Roch et al., 2023). For example, the OAK-D Camera, also known as the OpenCV AI Kit, is a smart camera that is great for indoor use. It can automatically process data files and use neural reasoning right inside the camera, without needing extra computer power from the robot. This means it can run neural network models without making the robot’s operating system work harder (Han et al., 2023).

In SLAM algorithms for robotics, scalability is a vital factor to keep in mind during the design of the system Middleware architecture. It enables rapid situational awareness over large areas, supports flexible dense metric-semantic SLAM in multi-robot systems, and facilitates fast map learning in unknown environments (Castro, 2021). This parameter needs to evaluate the algorithm’s capability to adjust to different mapping sizes and environmental conditions, particularly considering light emission, video, and/or image clarity. It should also provide versatility for various application needs, applicable to both indoor and outdoor scenarios (Laidlow et al., 2019; Zhang et al., 2023).

The ability of a SLAM algorithm to handle dynamic objects in the environment is an important consideration for robotics systems. This parameter assesses the algorithm’s ability to detect, track, and incorporate dynamic objects and moving obstacles into the mapping process (Lopez et al., 2020). It focuses on the algorithm’s capability to enable the robot to handle these objects effectively and respond quickly during the ongoing SLAM process (Wu et al., 2022). A robust dynamic environment should ensure the algorithm’s ability to adapt and respond in real-time applications. This is crucial for systems operating in environments where changes occur instantaneously, such as in interactive robotics applications (Li et al., 2018).

When choosing a SLAM algorithm for our project, it is important to observe whether it is open-source and has a community of active users. It is important because it makes it easier to customize and adapt the system according to our needs, benefiting from the experiences of the user community (Khoyani and Amini 2023; Xiao et al., 2019). Additionally, having community support ensures that the algorithm receives updates, bug fixes, and improvements. This enhances the reliability and longevity of the algorithm, making it better equipped to handle challenges during system implementation (Persson et al., 2023).

This parameter focuses on how a SLAM algorithm is represented and manages maps, allowing the researcher to determine its suitability for system hardware implementation. The evaluation includes the chosen method’s map representation, whether it is grid-based, feature-based, or point cloud, helping in assessing the efficiency of storing map information in the robotic system without encountering challenges (Persson et al., 2023; Acosta-Amaya et al., 2023). The selection of map representation influences memory usage and computational demands. It is a critical factor for robotic applications, especially those based on CNN and deep learning approaches (Duan et al., 2019).

In conclusion, we have summarized the preceding details in Table 2, offering a comprehensive overview of various V-SLAM algorithms. This table serves as a valuable resource for informed algorithm selection with comparative details for each method. It offers insights into the sensor capabilities, examining the types of sensors most effectively used by each algorithm and their role in facilitating algorithmic functionality. Moreover, the table underscores the potential application domains of the methods, empowering researchers to align their research objectives with suitable V-SLAM methodologies. The table also classifies algorithms based on their mapping scale distinguishing between small-scale (up to 100 m), medium-scale (up to 500 m), and large-scale (1 km and beyond) mapping capabilities (Tian et al., 2023b; Hong et al., 2021).

It also assesses their performance under varying illumination conditions, classifying algorithms based on their robustness, with categories ranging from the lowest, which represents (+) and to the highest which represents (+++++). Additionally, the table categorizes the algorithms based on their range of light intensity (RoLI), which reflects the robot’s ability to operate effectively in diverse lighting conditions, spanning from very dim to extremely bright. Moreover, the tolerance to directionality (T2D) category assesses the algorithm’s ability to function in environments with strong directional light sources, such as spotlights and windows. Collectively, these criteria collectively furnish a valuable resource for researchers seeking to pick the most fitting SLAM approach for their specific research endeavors.