Microsoft designed and released Kinect K1 (in 2010) and K2 (in 2014), which consist of one IR emitter, one IR camera, and one RGB camera, to gain depth and color images. To strongly enter the gaming industry and provide human connections as a part of Xbox consoles, targeted the companies that designed the Kinect, which has lately been used as a markerless, affordable, and portable MoCap sensor (Bilesan et al., 2019). Kinect can perform PE using its built-in depth sensor and the associated software development kits (SDKs). Many studies explored the accuracy of the two Kinects in real-time skeletal tracking during human movements (Ganea et al., 2014; Kurillo et al., 2022). However, according to (Wang et al., 2015), regarding MoCap selection as a gold standard, the K2 generation is beneficial in terms of joint estimation accuracy and more robust in handling track occlusion and body rotation. In another study, a Vicon MoCap was compared with Kinect in gait analysis, as participants walked and jogged on a treadmill. The results show the validity of the Kinect in the sagittal plane for stride timing in kinematic analysis (Pfister et al., 2014).

Depending on the coaches’ purpose, estimated pose through the Kinect somatosensory camera can be sufficient in physical activity recognition, such as qigong movement detection (Fan et al., 2024). Or by requiring more complex analysis processes, their implementation with other codes or devices, such as 3DSMAX, can invent novel methods for monitoring the sports training seasons (Shi, 2014). Additionally, regarding simulating and estimating the body joints and limbs, the skeletal body structure, Asteriadis et al. used Kinect in body estimation in 3D by minimizing the noise (Asteriadis et al., 2013).

The Kinect device alone lacks accuracy in defining orthogonal axes unrelated to joint centers, limiting its ability to measure internal or external angular rotations and displacement. Adding depth sensors, video cameras, or multiple devices to make a MoCap-based Kinect can improve the tracking process and angular measurement accuracy (Clark et al., 2012; Napoli et al., 2017).

The ZED Camera by StereoLabs, released in 2015, was followed by the next-generation, including ZED2 (2019) and ZED2i (2021), which were enhanced with depth sensing and AI features (Sarıalioğlu et al., 2024). ZED cameras are equipped with two high-resolution lenses to capture in-depth information, making them suitable commercial vision hardware for integrating open pose libraries in kinematic and biomechanical studies (Avogaro et al., 2023). A comparative analysis of the ZED2 camera with other commercial vision systems, including Azure Kinect and RealSense D455, utilizing their respective SDKs, shows higher performance in human pose tracking at distances exceeding 3 m. This advantage aids PE in larger spaces or dynamic environments, such as field tests (Ramasubramanian et al., 2024). Another suggests that Nuitrack and MediaPipe software integrate with ZED2i to perform better in upper and lower extremity features for pose tracking (Aharony et al., 2024). Combining ZEDmini with a Virtual Reality (VR) headset and an augmented reality (AR) prototype resulted in a device capable of recognizing body formation and sports equipment such as a tennis racket and basketball hoop. This novelty aids players with low vision to enhance their ability to identify the players and objects (Lee et al., 2024). The advantages of ZED cameras as hardware for exercise monitoring, assessing gait impairment in Neurodegenerative disorders, and joint angle measurement include their ease of use, affordability, and ability to deliver reliable results in both 2D and 3D kinematic analyses of exercises, such as squats (Aliprandi et al., 2023; Zanela et al., 2022).

Intel RealSense cameras combine high-resolution RGB imaging with depth perception, offering models like the D400 series for general depth sensing, the L515 for precise LiDAR tasks, and the T265 for motion tracking with simultaneous localization and mapping (SLAM) capabilities (Moghari et al., 2024; Tsykunov et al., 2020). By capturing depth maps and RGB frames, they integrate seamlessly with pose-estimation software framework tools, mapping 2D keypoints to real-world coordinates. Their lightweight design, robust SDK, and real-time performance make them versatile for indoor and outdoor applications, outperforming infrared-based systems like Kinect in varied lighting conditions (Maddipatla et al., 2023).

An Intel RealSense camera is used in head motion tracking using custom software while participants walk indoors and outdoors. They walked in both slow and fast modes, and their head motion tracking was compared with that of a perambulator containing inertial measurement units (IMUs) and a distance counter, serving as the gold standard. The accurate results of this gait study showed that while the participants walked slowly (Hausamann et al., 2021). Another study compared the Intel RealSense camera to the Kinect in slow walking gait analysis. The results indicated that while the Intel RealSense D435 is valid for measuring certain spatiotemporal variables, the Kinect proved more reliable for collecting skeletal data due to its more robust RGB-D camera (Mejia-Trujillo et al., 2019). MediaPipe landmarks and two Intel RealSense cameras were used to estimate angular parameters in five exercises, including squats, knee, ankle, hip extension, and shoulder elevation, compared with OptiTrack as the reference. The results showed an error of less than 20 degrees, demonstrating an adequate estimation quality. Still, further improvements in both software and hardware can enhance the system’s accuracy (Pilla-Barroso et al., 2024). These generations of cameras are increasingly being used in sports performance analysis through integration with AI-based markerless MoCap software. Intel RealSense depth cameras serve as sufficient registration tools in markerless PE, especially for closer range (under 3 m) (Ramasubramanian et al., 2024; Jacobsson et al., 2023).

OpenPose is an open-source library developed by Carnegie Mellon University for 2D PE of the skeletal structure by detecting 25 keypoints on the body (Hidalgo et al., 2019). A real-time detection system may focus on specific angles between joints, the center of mass, or a point of displacement (Cronin et al., 2024; Needham et al., 2021). To analyze the running performance and timing, OpenPose evaluated it with the Coco dataset regarding human PE to recognize hurdles athletes’ wrists, ankles, hips, and knees (Jafarzadeh et al., 2021; Sharma et al., 2022). The Coco dataset recognizes 18 pre-trained landmarks with x, y, and v values, where x and y indicate coordinates, and v shows the visibility of the landmarks (Duan et al., 2023).

PE relies on video quality and frame rates, meaning that weather conditions and lighting impact the accuracy of ML systems. A study evaluating the AI accuracy of OpenPose as a PE model under various environmental and lighting conditions, while participants performed stretching exercises, found that the accuracy rates depend on the environment and setup context (Song and Chen, 2024).

In a study using 3D motion analysis as the reference standard, an OpenPose-based motion analysis was used to compare knee valgus angular data during the drop vertical jump test. The strong correlation among these data indicated a sufficient ML accuracy level of the PE models in measuring angular movements (Ino et al., 2024). In baseball, analyzing a hitter’s limb angles and hip distance aids in swing performance (Li et al., 2021). AI-driven gesture recognition supports biomechanical analysis in training techniques and game strategy evaluation. This approach was a key objective in a study analyzing tennis player performance through skeletal PE (Wu et al., 2023).

Furthermore, PE models’ integration with wearable sensors or MoCap can facilitate more complex processes through Taekwondo kinematic analysis through a PE model, 8 body joint angles in players’ upper and lower body registered by Contemplas cameras, and the Vicon system. OpenPose was used to process recordings from Contemplas cameras to extract 2D human skeleton data, then triangulated to obtain 3D joint angles to compare world-class and master-class players’ kicking performance (Fukushima et al., 2024). Plus, optimal and kinetic analysis of fouling and shooting in basketball player performance was conducted using video-based (2D) methods with a lightweight deep learning (DL) architecture, where OpenPose was used as the PE model (Xu, 2024). This open-source framework is being used in athletes’ 3D estimating with integration with sensors or the Internet of Things (IoT) in various sports such as gymnastics (Ren et al., 2025).

MediaPipe is an open-source framework developed by Google for 2D human PE (Pham et al., 2022). It enhances AI applications by detecting keypoints on the body, face, and hands in images or live video streams to offer gesture recognition, motion analysis, and interactive systems (Saini, 2024). MediaPipe is a broader framework that includes BlazePose and other ML solutions, where BlazePose specializes in detecting 33 keypoints (Wang, 2024).

An important finding shows that when assessing PE quality using MediaPipe from a frontal and lateral view, an error margin of less than 25% of the range of motion is optimal, with specific tolerances: under 20 degrees for movements greater than 90°, around 10 degrees for movements up to 40°, and below 10° for static angles (Pilla-Barroso et al., 2024).

The validity of MediaPipe BlazePose in measuring joint angles compared with the IMU-based MoCap in 2D image analysis, and the results showed that the difference in accuracy is within 10% (Pau et al., 2021). In addition, an ML tool using MediaPipe simulated kinematic points, assuming shoulder, knee, hip, and ankle joints for artistic swimming analysis. Regarding the 2D kinematic simulation frames, the leg deviation angles were measured during execution and body position in preparation to aid officiating systems and performance monitoring (Edriss et al., 2024b).

In a video performance analysis using an ML tool with MediaPipe, the knee angles of advanced and beginner pickleball players were measured during dink shots. This kinematic analysis assessed the differences in knee bending and body flexion between the two groups, suggesting an appropriate positioning formation and preparation for dinking (Edriss et al., 2025).

An application using MediaPipe, compared with Kinect, was used to measure knee angles for athletes under a motion test. The results showed the higher capacity of the MediaPipe framework in analyzing ACL injury risk by calculating lower limb kinematics through a smartphone (Babouras et al., 2024). However, a limitation of the MediaPipe source is the lack of ability to recognize multiple people in a frame (Dong and Yan, 2024).

AlphaPose is an open-source tool that provides PE by detecting 25 keypoints. It was developed by the Chinese University of Hong Kong (Fang et al., 2023). Like OpenPose, AlphaPose provides a high keypoint detection rate and is highly interchangeable with ML models for biomechanical PE parameters like ground reaction forces (Mundt et al., 2023).

This framework is used in various ways, including gait and player kinematic analyses. For example, in gait analysis, participants walked at different speeds, and their estimated poses were captured using AlphaPose (Wei et al., 2023). Additionally, AlphaPose captured accurate skeleton PE of badminton players to classify their movements. These results achieved an 80% prediction accuracy, effectively analyzing the sequential nature of badminton actions (Liang and Nyamasvisva, 2023).

Several researchers are working on PE framework development through the DL tool, such as YOLO (Yang et al., 2024; Zhao et al., 2024). Although the validity of AlphaPose is sufficient, its accuracy compared to the data of other frameworks is considered. For example, a study compared MediaPipe BlazePose and AlphaPose with Vicon MoCap as the reference standard in gait joint kinematics analysis. This study emphasized that MediaPipe shows a lower mean square error (RMSE), indicating more accurate joint kinematic measurements. However, AlphaPose showed greater variability with a higher range of motion and RMSE (Hulleck et al., 2023).

DensePose is a PE model developed by Facebook AI Research that maps human pixels in an image to the body’s 3D surface (Gu et al., 2022; Lovanshi and Tiwari, 2022; Mahajan et al., 2023). The advantages of DensePose, compared to other frameworks, are its perfect ability for multi-person recognition and its capability to estimate the body by keypoints or mesh (Güler et al., 2018; Guler and Kokkinos, 2019; Zhang et al., 2021). DensePose is effective for applications in augmented reality, virtual try-on, and immersive human-body interaction scenarios (Islam et al., 2024; Zhu and Song, 2021). Athletes’ biomechanical and kinematic analyses can be conducted by designing a 2D PE from videos using DensePose, which is pre-trained on the COCO keypoint dataset. For monocular 3D PE, MotionAGFormer, pre-trained on the Human3.6M dataset, can be utilized. DensePose requires more technical skills to set up due to its powerful GPU resources and custom code for mesh mapping. Studies indicate that joint angular measurements obtained through DensePose also provide reliable data (Hellstén et al., 2022; Suzuki et al., 2024).

Figure 1i illustrates the shape of the vision sensors, and Figure 1ii demonstrates the human skeletal structure as detected by the PE models.

Table 1 shows a brief comparison between hardware and software to achieve a clearer idea of their strength and pros.