Emerging trends in gait recognition based on deep learning: a survey

Advantages of gait recognition technology

When compared to other biometric identification systems Gait identification system has several advantages.

Long-distance capture: Gait recognition technology can accurately capture walking pattern of an individual from a distance, making it suitable for surveillance and large-scale video analysis (Su, Zhao & Li, 2021; Zhang et al., 2022; Mu et al., 2020; Han et al., 2022; Jin et al., 2022).

Non-cooperation requirement: Unlike other biometric recognition technologies that require active participation from the individual, gait recognition can be applied unobtrusively, without the cooperation of the person (Su, Zhao & Li, 2021; Zhang et al., 2022; Mu et al., 2020).

Difficulty in imitation: Gait recognition is difficult to fake or manipulate, making it a reliable means of identification (Cai et al., 2021; Aung & Pluempitiwiriyawej, 2020). Additionally, gait recognition has proven useful in crime investigation, forensic identification, and social security applications (Liu et al., 2021; Lin et al., 2021).

Usefulness of gait recognition technology

The utilization of gait recognition technology has transformed a number of industries, including robotics, sports, law enforcement, and healthcare. This modern technology has unmatched potential for improving forensic investigations, reviving cold case inquiries, and validating testimonies and alibis. Gait analysis is essential for fall detection, injury prevention, and rehabilitation monitoring in medical settings. It also helps athletes and sportspeople perform better and avoid injuries. Its applications include wearable technology, biometric identification, human-computer interfaces, vehicle safety, and robotics, which makes it a multipurpose tool with enormous promise in a variety of fields. This in-depth study explored the various ways in which gait recognition technology is reshaping industries and advancing human endeavors.

Cross-checking alibis and testimonies (Liu et al., 2021; Lin et al., 2021)

Utilizing spatiotemporal methods, gait recognition serves as a crucial tool for verifying or contradicting alibis and testimonies. Surveillance footage that captures unique gait patterns provides evidence of regarding the presence of an individual at a crime scene. By comparing these patterns with claimed alibis or witness descriptions, it verifies or contradicts the presence of individuals at specific locations and times. This objective evidence enhances the reliability of alibi verification and witness testimonies, strengthens the investigative process and contributes to the accuracy and fairness of legal proceedings.

Criminal profiling (Liu et al., 2021; Lin et al., 2021)

Gait analysis in criminal profiling harnesses the unique walking styles of individuals, akin to fingerprints or DNA, to create suspect profiles. This process involves scrutinizing factors like stride length, speed, posture, and foot placement. By scrutinizing surveillance footage, law enforcement can extract valuable information on the gait patterns of a suspect, aiding in narrowing down the potential leads. In cases where traditional identification methods are inconclusive, gait analysis offers an additional means of pinpointing suspects by comparing observed patterns with databases. This technique not only facilitates the identification of potential suspects, but also aids in prioritizing investigative resources more effectively, increasing the likelihood of apprehending perpetrators.

Forensic evidence (Liu et al., 2021; Lin et al., 2021)

Gait recognition technology enhances forensic investigations by analyzing distinctive walking patterns of individuals captured in surveillance footage. By correlating these patterns with specific locations and times, the presence or movement of individuals during critical events can be confirmed. These empirical data reinforce other case materials, such as DNA analysis or fingerprinting, by corroborating timelines and establishing connections between individuals and crime scenes. Moreover, gait recognition fills the gaps in traditional forensic evidence, particularly in cases of obscured visual identification. Overall, it strengthens the integrity of legal proceedings and contributes to more robust outcomes in the criminal justice system.

Investigation of unsolved cases (Liu et al., 2021; Lin et al., 2021)

Gait recognition technology is instrumental in reinvigorating investigations of unsolved cases, particularly cold criminal cases that have stalled traditional methods. By analyzing unique walking patterns, it introduces a new approach that offers new leads and avenues for investigation. It allows investigators to revisit existing evidence, focusing on the movement of individuals and behaviors, potentially uncovering overlooked details or connections. Moreover, gait recognition overcomes the limitations of traditional methods by providing alternative means of identification, even in cases of obscured faces.

Enhancement of witness testimonies (Liu et al., 2021; Lin et al., 2021)

Corroborating witness descriptions: By analyzing surveillance footage to match the gait pattern of an individual with witness-provided physical descriptions, gait recognition adds credibility to the testimonies.

Verifying movements: Gait analysis ensures alignment between witnessed movements and those observed in surveillance footage, thereby substantiating the accuracy of testimonies regarding suspect actions.

Challenging inconsistencies: Gait recognition identifies disparities between witness accounts and observed gait patterns, prompting critical examination of testimonies and potential further investigation.

Providing objective evidence: Gait analysis furnishes impartial evidence grounded in unique walking patterns, augmenting the reliability of testimonies with scientific rigor.

Healthcare and rehabilitation (Sethi, Bharti & Prakash, 2022; Connor & Ross, 2018; Kumar et al., 2021)

Rehabilitation monitoring: To track and evaluate the progress of patients during rehabilitation programs, physical therapy settings make extensive use of gait technology. Through consistent gait analysis, therapists can objectively monitor the advancements or modifications over time. Therapists can customize treatments and therapy regimens according to the unique requirements and advancements of each patient using this data-driven approach. For example, therapists can modify exercises or therapies to address areas of weakness or imbalance, if gait analysis indicates abnormalities or uneven movement patterns. Ultimately, this customized strategy improves rehabilitation programs efficacy and aids patients in achieving the best possible recovery.

Fall detection: Defects in walking patterns can also be a sign of an elevated fall risk, especially in the senior population, where gait analysis is useful. Before a fall occurs, gait recognition technology can detect people who may be at risk of falling by examining minor changes or departures from typical gait patterns. By taking a proactive stance, healthcare professionals can reduce the likelihood of falls by implementing interventions or preventive measures, such as proposing assistive technology, making environmental modifications to eliminate dangers, or prescribing exercises to enhance balance and stability. By encouraging safety and independence, the early identification of fall risk through gait analysis of an individual improves the overall quality of life and reduces the likelihood of accidents and hospital stays.

Sports and athletics (Sethi, Bharti & Prakash, 2022; Marsico & Mecca, 2019; Kumar et al., 2021)

Performance optimization: Athletes use gait analysis to refine their walking or running form and increase effectiveness. Gait recognition technology delves into the biomechanical elements that examine the gait pattern of an athlete under various factors such as posture, foot placement, cadence, and stride length. With these data, athletes can modify their technique to optimize energy use, minimize extraneous motions, and enhance overall efficiency. For instance, an athlete overstriding or excessive lateral movement, may be identified by gait analysis as inefficiencies in their running form. This can be corrected using focused training and technical modifications. Athletes can enhance their athletic prowess and perform better in competitions by fine-tuning their walking patterns.

Injury prevention: Coaches and sports medicine professionals use gait technology to identify biomechanical issues that may contribute to injuries, allowing for targeted interventions.

Biomechanics research (Liu et al., 2021; Lin et al., 2021; Sethi, Bharti & Prakash, 2022; Kumar et al., 2021)

Human movement studies: Researchers use gait analysis to understand the mechanics of human movement, which is valuable for designing prosthetics, orthotics, and ergonomic products.

Clinical studies: Gait technology assists in clinical studies related to neurological disorders, musculoskeletal conditions, and other health-related research.

Security and surveillance (Liu et al., 2021; Lin et al., 2021; Sethi, Bharti & Prakash, 2022)

Biometric identification: Gait analysis leverages distinctive walking patterns as a biometric identifier for authentication and identification. Similar to fingerprints or facial features, gait patterns are also unique and can be used as reliable markers for distinguishing one individual from another.

Human-computer interaction (Sethi, Bharti & Prakash, 2022; Sun, Su & Fan, 2022)

Natural interaction: Integrating gait analysis into gesture recognition systems enables fluid, and instinctive interactions without traditional input devices. Users control virtual environments with subtle movements, and enhance immersion by eliminating the need for controllers.

VR immersion: Gait recognition allows users to navigate virtual worlds naturally using body movements to interact with digital content. This deepens immersion as users can walk, run, and gesture to manipulate objects and explore virtual spaces.

Gesture-based gameplay: Gait technology introduces gesture-based gameplay, translating walking patterns into in-game actions. Users can jump, crouch, or interact with objects by performing specific gestures while walking, adding physicality, and engaging in gaming.

Accessibility: Gait-based gestures enhance accessibility by offering alternative input methods to users with mobility limitations. By leveraging natural body movements, these systems are inclusive, allowing a wider range of users to participate in VR and gaming activities equally.

Wearable technology (Sethi, Bharti & Prakash, 2022; Marsico & Mecca, 2019; Connor & Ross, 2018)

Fitness tracking: Users who wear wearable technology with gait analysis capabilities can obtain comprehensive insights into their jogging or walking habits. These gadgets provide useful information for evaluating the overall fitness and health of an individual. Users can set fitness goals, monitor their progress over time, and modify their routines, as necessary.

Assistive devices: Gait recognition technology in assistive devices like smart shoes or insoles, offers immediate feedback and aids for people with mobility impairments. These devices analyze gait patterns to deliver personalized assistance, and enhance stability, balance, and mobility. For instance, smart shoes with gait sensors can detect irregular walking patterns and provide feedback to correct the posture or stride.

Automotive safety (Sethi, Bharti & Prakash, 2022; Marsico & Mecca, 2019)

Driver monitoring systems utilize gait analysis to evaluate driver behavior and conditions, identifying slight alterations in walking patterns that could signify fatigue or alertness. This approach enables the ongoing monitoring of driver alertness, facilitating prompt action to avert accidents. Gait analysis is particularly adept at detecting fatigue indicators, which is a notable hazard for road safety. Immediate alerts and interventions can be provided to drivers, suggesting breaks or activating automated driving modes to ensure safe vehicle operation until alertness is restored.

Robotics (Sethi, Bharti & Prakash, 2022; Sun, Su & Fan, 2022)

Humanoid robots: Gait technology can be integrated into humanoid robot designs to facilitate lifelike walking patterns and improve the overall interaction between humans and robots. Anthropomorphic robots are machines that act more like humans. This helps people to better utilize and accept them. This also makes people feel more connected to machines, which is important for rehabilitation. Machines with human-like traits can communicate seamlessly with users. By incorporating findings from gait research, robot designers can replicate human movements, such as stride length, cadence, and posture, thereby enhancing the authenticity of the walking motion. This makes it possible for humanoid robots to engage with people in a more natural, move more adaptably through different contexts, and offer individualized physical assistance, especially in healthcare settings. Robots with gait recognition technology can predict and react to human motions by analyzing human gait patterns in real-time. This capability broadens the range of applications of robots and fosters cooperation between people and robots under a variety of circumstances.

Appearance-based methods

Silhouette-based gait analysis: (Han & Bhanu, 2006; Lam, Cheung & Liu, 2011; Tao et al., 2007). This method extracts the silhouette of a person while walking and analyses the shape and motion of the silhouette. Features such as the aspect ratio, area, and height were used for recognition.

Image and video-based gait recognition: (Işık & Ekenel, 2021; Wang et al., 2010; Han et al., 2022; Mowbray & Nixon, 2003). These methods use images or videos of a person walking, and then extract features from these images or frames. This includes traditional computer vision techniques and deep learning methods.

Model-based methods

Dynamic time warping (DTW): (Zhang et al., 2019; Bashir, Xiang & Gong, 2010). It is appropriate for gait identification because it calculates the similarity between two temporal sequences. The algorithm aligns the sequences and calculates the distance metric.

Hidden Markov models (HMMs): (Kale et al., 2004; Chen, Wu & Li, 2020; Zhang et al., 2021) HMMs model the temporal dynamics of gait patterns. Each gait sequence is represented as a sequence of states, and the model can be trained to recognize individuals based on these sequences.

Principal component analysis (PCA): (Zhang et al., 2019; Ariyanto & Nixon, 2012; Bobick & Johnson, 2001; Cunado, Nixon & Carter, 2003) PCA was used for dimensionality reduction of the gait data. It can be applied to feature extraction to reduce the complexity of the gait patterns.

Sensor-based methods (Sethi, Bharti & Prakash 2022)

Inertial sensors: These methods use accelerometers and gyroscopes to capture motion data while walking. Machine learning algorithms can then process these data for gait recognition.

Pressure sensors: Pressure sensors embedded in the floor or insoles of shoes can capture foot pressure patterns during walking that are unique to individuals.

Deep learning methods (Chen, Wu & Li, 2020; Zhang, Wang & Li, 2021; Ghosh, 2022)

Convolutional neural networks (CNNs): CNNs can be used for gait recognition by processing images and video data. They automatically learn relevant features from raw image data.

Recurrent neural networks (RNNs): play a role in gait identification problems because they can be used to model temporal dependencies in gait sequences.

Siamese networks: By learning to distinguish gait pattern among various people, these networks, which are made for one-shot learning tasks, can be utilized for gait recognition.

Fusion methods (Zhang et al., 2021; Zhang et al., 2022; Cai et al., 2021; Cosma & Radoi, 2020; Teepe et al., 2021; Ghosh, 2022)

Multimodal fusion: This approach combines information from multiple sources such as video, depth sensors, and other biometric modalities (such as face or voice recognition) to improve gait recognition accuracy.

Score-level fusion: Different recognition methods, such as appearance-based and model-based, can produce scores for each individual. These scores can be fused to make a final recognition decision.

3D gait recognition (Sethi, Bharti & Prakash 2022)

Depth cameras: Depth cameras like Microsoft Kinect or LiDAR sensors capture three-dimensional information of a person during walking, which can be used for gait recognition.

Spatiotemporal method (Su, Zhao & Li, 2021; Khan, Farid & Grzegorzek, 2023)

Spatiotemporal template: This method captures both spatial and temporal information regarding the movement of a person to create templates for gait recognition.

Gait energy image (GEI): GEI is a representation that combines multiple frames of silhouette into a single image, thereby emphasizing the energy distribution of the gait cycle.

• In relation to the gait recognition research in previous studies, there was a preference for model-based approaches initially (Zhang et al., 2021; Kale et al., 2004; Wagg & Nixon, 2004; Yam, Nixon & Carter, 2004; Yamauchi, Bhanu & Saito, 2010; Liao et al., 2017; Ariyanto & Nixon, 2012; Bobick & Johnson, 2001; Cunado, Nixon & Carter, 2003). However, there has been a shift towards the appearance-based methods owing to their increased popularity (Işık & Ekenel, 2021; Wang et al., 2010; Han & Bhanu, 2006; Makihara et al., 2006; Tao et al., 2007; Lam, Cheung & Liu, 2011; Wang et al., 2012; Bashir, Xiang & Gong, 2010; Mowbray & Nixon, 2003; Liu & Sarkar, 2006; Zhang et al., 2019). This shift has been driven by the ability of appearance-based methods to overcome the limitations associated with the former, such as high computational complexity, limited robustness to variations, sensitivity to environmental conditions, privacy concerns, and complex implementation.

• Appearance-based approaches offer advantages, such as reduced data requirements, lower computational complexity, robustness to variations, simplicity of implementation, and faster processing speed, making them practical for real-world applications. However, these methods have limitations, including sensitivity to environmental conditions, limited discriminative power, susceptibility to clothing variations, and privacy concerns.

• Deep learning has revolutionized gait identification by addressing these challenges through automatic feature learning, robustness to variations, transfer learning, and temporal modelling. Despite its impressive accuracy, deep learning requires a substantial amount of labelled data and can be computationally intensive, leading to ongoing research for improvement.

• Gait recognition researchers commonly employ a combination of techniques, including deep learning and spatiotemporal methods to enhance the accuracy and robustness. Sensor-based models require wearable devices, whereas appearance-based and 3D gait recognition primarily rely on visual data, with method choice influenced by the application, data accessibility, and accuracy requirements.

Researchers continue to explore and integrate various techniques to advance gait recognition. In this article, we mainly focused on the deep learning techniques and their combinations.

Search strategy analysis

This research article draws from various databases, including SCOPUS, IEEE Xplore, Google Scholar, Web of Science, Springer, and Science Direct, covering top-indexed journals and high-impact research articles in the field from 2020 to 2023. The primary aim of the survey was to provide a comprehensive analysis of the latest developments in gait recognition for person identification, thus setting the timeframe since 2020. Relevant articles containing the keywords “gait recognition using deep learning” and “person identification” were identified without the need for additional refinement using wildcards. The focus is solely on identifying individuals based on gait recognition techniques without requiring cooperation. Consequently, the consideration of wearable devices, such as smart devices, and non-wearable devices, like sensor-based models, is excluded. Table 1 represents the keywords used to identify the research articles in their database.

Inclusion and exclusion criteria

The inclusion and exclusion criteria while searching the research articles in their respective databases and PRISMA analysis diagrams were also framed. Figure 2 shows a PRISMA analysis diagram for the selection of articles. Table 2 shows the inclusion criteria for fine-tuning the selection process and the exclusion criteria to eliminate irrelevant topics. Of the 1,391 records, 22 were included in the qualitative synthesis after screening, eligibility assessment, and removal of duplicates and exclusions, resulting in a final selection of 22 articles.

Selection of research data

The selection of research data can be explained using a step-by-step approach, which is also represented in Fig. 3.

Stage 1: Collect research articles from various databases such as SCOPUS, IEEE Xplore, Google Scholar, Web of Science, Springer, and Science Direct using the keywords “Gait recognition using deep learning” and “Person Identification”.

Stage 2: Filter articles published between 2020 and 2023, excluding wearable and non-wearable sensor-based hybrid models.

Stage 3: Select articles based on title/abstract relevance and eliminate irrelevant articles.

Stage 4: Articles with incomplete data, deviation from the topic, or improper content were excluded.

Stage 5: Remove reviews/survey articles and include highly related abstracts and full-text articles.

CNN based models

A CNN is a deep learning model that specializes in handling visual data, such as images and videos, with exceptional effectiveness. This section provides an overview of several models based on CNNs that are used for human gait recognition, which is the process of recognizing people based on the way they walk. The distinct methodologies of each subsection are explained, including the disentangled framework of GA-ICDNet (Chai et al., 2021), the incremental learning technique of iLGaCo (Mu et al., 2020), and the many ways in which CNN, LSTM, and Mask R-CNN are combined to process and extract features. These models enhance the gait identification performance on a variety of datasets and in practical applications by employing distinct loss functions and feature extraction methods.

GCN based models

Neural networks with graph convolutional networks (GCNs) have been specifically engineered to handle graph-structured input. In gait recognition, GCNs play a key role in analyzing walking patterns expressed as graphs, where nodes correspond to component features (e.g., joint positions) and edges denote the relationships between these features. By aggregating information from neighboring nodes and edges, GCNs enable the extraction of meaningful gait features that capture both the spatial and temporal characteristics of walking patterns. In this section, research has explored gait recognition from UAVs at altitudes of 10, 20, and 30 m using innovative models: one integrates graph convolution with traditional CNNs, another combines skeleton poses with a GCN, and a third merges CNN, LSTM, and GCN for hybrid silhouette-skeleton gait representation. These approaches enhance feature extraction and spatiotemporal representation, and improve gait recognition accuracy.

Fully connected network model

Zhang et al. (2021) introduced a model which is composed of three primary components: the initial component is an Observation Function Approximating module or OFA, followed by a Koopman Matrix Memory or KMM, and finally, the system includes a Discriminative Feature Extractor module or DFE. The OFA (Kingma & Welling, 2013) fed the input image to an encoder, and the encoder used a non-linear deep network transformation to convert the original input data into Koopman space. To prevent the code in the Koopman space from converging on outliers, such as zeros, and to maintain the majority of useful information in the original images, a decoder is used. Koopman Matrix Memory (KMM) (Rowley et al., 2009; Williams, Kevrekidis & Rowley, 2015) can be used for system state analysis by expressing a nonlinear dynamical system in a linear space using the Koopman operator. Finally, the committed Koopman matrix was transformed into a fresh feature within a discriminative space using a simple fully connected network. The DFE module uses softmax loss and triplet loss with hard mining (Hermans, Beyer & Leibe, 2017) as its two identity identification methods.

GaitSet employed by Han et al. (2022), is composed of a combination of different losses and has two stages: the training stage and the testing stage. Alignment and network training are the two key stages of the training process. The testing procedure consisted of three steps: feature extraction, feature alignment, and gallery search. Similar to the training stage, the silhouettes were first set to a predetermined size. The trained network was then probed with walking silhouettes to extract gait features. For verification, the closest neighbor was measured against a predefined threshold. The identification of the nearest neighbor is the outcome of the recognition if the distance is less than the threshold. A-Softmax loss (Liu et al., 2017) and triplet loss were used to train the suggested model simultaneously. While the A-Softmax loss applies a decrease in intra-class distance and an increase in inter-class distance to capture discriminative traits, the triplet loss of Hermans, Beyer & Leibe (2017) imposes an angular margin to extract separable features. A batch-normalization layer (Luo et al., 2019) was used to enable the training process once the features were extracted.

Temporal part-based GaitPart model

In the work published by Fan et al. (2020), the integration of FPFE, which stands for Frame-level Part Feature Extractor, and TFA, which stands for the Temporal Feature Aggregator was carried out. The FPFE extracts spatial features that consider specific body parts for each frame. This is achieved through a convolutional network comprising three blocks, where each block comprises a pair of FConv or Focal Convolution Layer. The feature map undergoes horizontal division into n segments using the Horizontal Pooling (HP) module, which is designed to capture distinctive part-specific features of human body segments. The work of modeling the short-range spatiotemporal properties of each corresponding body segment was divided across n parallel MCMs within the TFA. The loss function employed here is a Separate TripletLoss.

Hidden markov model

The HMM is an easy-to-understand model that can effectively categorize data and appropriately describe a dynamic time series. In the work published by Chen, Wu & Li (2020), a combined HMM and GFHI or Hidden Markov Model and Gait Optical Flow History Image were used. The Motion History Image (MHI) and the optical flow grayscale image serve as the foundation for the GFHI (Horn & Schunck, 1980). The optical flow can be used to calculate the local movement speed of the target. Because the MHI is a grayscale image containing temporal information, the pixel value for the most recent motion is higher. The network uses Hu Invariant Moment Feature Extraction (Hu, 1962) to identify the local details in the image.

GI-ReID model

The main goal of Jin et al. (2022) was to recognize the same individual repeatedly in different contexts. CC-ReID represents cloth-changing person re-identification whereas the GI-ReID system makes every effort to fully utilize the characteristic human gait to resolve the cloth-changing challenge of ReID by using solely one image. The GI-ReID system utilizes a dual-stream framework that consists of an image ReID-Stream and an extra stream designed for gait recognition, referred to as Gait-Stream. These two streams, ReID-Stream and Gait-Stream, were trained simultaneously while adhering to a high-level semantic consistency (SC) constraint. The learning of cloth-independent feature sets is regulated by the gait characteristics in the ReID-Stream. The gait stream consisted of two components: a pre-trained gait recognition network, GaitSet (Chao et al., 2019), and a gait sequence prediction module. The purpose of the GSP or gait sequence prediction module is to enhance the gait data. The training of ReID-Stream is then guided by GaitSet using augmented gait features that are both independent of clothing and discriminative in terms of motion signals.

Combination of faster R-CNN and RNN

Implementing a modified Faster R-CNN architecture, the proposed method by Ghosh (2022) first determines whether a human is present in any of the frames of the walking activity video. Instead of the five convolutional layers of the classic Faster R-CNNs, nine distinct convolutional layers were used in this enhanced version. The aspect ratios of the anchor boxes were modified to accommodate walkers with various widths and heights. The RPN uses a tiny network to create region suggestions. The feature matrix resulting from the final convolutional layer is slid over by using a single window. A 256-D network was mapped to each sliding window to acquire spatial information for each site. The number of region proposals projected for each location depended on the number of anchors in a single sliding window. Nine anchor boxes were used for each site in this work, which employing three scales and three aspect ratios. The RPN creates region proposals with five-tuple values (index, x, y, w, h) of different sizes. The RoI or Region of Interest pooling layer projects the feature matrix generated from the final convolutional layer onto it. Among the positive suggestions labeled by the RPN, this layer selects 100 RoI proposals that exhibit higher Intersection over Union (IoU) values for this specific inquiry. Furthermore, the RoI pooling layer employs max-pooling on the feature matrix corresponding to each proposal. The proposed architecture simultaneously accomplishes two tasks: it assesses whether a pedestrian is present in the frame and it establishes a bounding box around the detected individual. Every region of interest (RoI) obtained from the RoI pooling layer undergoes individual processing by the fully connected (FC) layers. The output layer employs the softmax function to forecast the existence of a human, and a regressor to generate coordinates for a rectangular box enclosing the identified individual. The detection network calculated the classification and regression losses in the output layer. These losses must be calculated to identify the presence of a person and establish a bounding box.

OU-MVLP: Takemura et al. (2018)

Data source: A science museum protracted video-based gait analysis presentation yielded the data. All participants willingly consented to participate in the study.

Dataset size and diversity: The dataset encompasses 10,307 individuals, including 5,114 males and 5,193 females, covering a broad age spectrum from 2 to 87 years.

Multiple viewing angles: Gait images were captured from 14 distinct viewing angles, spanning from 0° to 90° and 180° to 270°, providing comprehensive coverage for gait analysis.

Image quality and frame rate: Image resolution: 1,280 × 980 pixels and Frame rate: 25 frames/seconds.

Camera configuration: Seven network cameras were strategically placed at 15-degree intervals along a quarter-circle path, with their centers coinciding with the walking area. This path had a radius of approximately 8 m and a height of approximately 5 m, enabling varied gait perspectives of the subjects.

OU-LP-Bag: Makihara et al. (2017)

This dataset was generated during a hands-on exhibition on video-based gait analysis at the Miraikan Science Museum, where visitors electronically provided consent for research. It included 62,528 subjects with ages spanning from 2 to 95 years. The camera setup ensured clear and detailed images, which allowed for comprehensive gait analysis with an image resolution of 1,280 × 980 pixels and a frame rate of 25 frames/second.

OU-LP-Bag β: Makihara et al. (2017)

Training set: This includes 2,068 sequences representing 1,034 individuals, with each person having two sequences, one with and one without carried objects.

Probe set: Comprising 1,036 subjects, this set features individuals who carry objects that are distinct from the training set.

Gallery set: The gallery set mirrors the probe set subjects, but shows them without carrying objects.

Front-view gait: Zhang et al. (2019)

Database creation: This database was assembled during the years 2017 and 2018 to support research in the field of gait recognition.

Privacy enhancement: To protect subject privacy, a modified version called front-view gait (FGV)-B was created, which blur facial features to a level where even advanced face recognition algorithms fail to identify individuals.

Challenging frontal view: Frontal-view walking is considered more challenging in gait recognition because it provides limited gait cues compared with other view angles.

Diverse conditions: FVG-B encompasses diverse conditions, including variations in walking speed, carried items, and clothing, all captured from a frontal perspective.

Database details: FVG-B included frontal walking videos from 226 subjects, with 12 subjects being recorded in both 2017 and 2018, resulting in a total of 2,856 videos. A tripod at a height of 1.5 m was used to hold either a GoPro Hero 5 or Logitech C920 Pro webcam to record the movies. They had a resolution of 1,080 × 1,920 pixels and an average duration of 10 s.

USF: Sarkar et al. (2005)

The USF Human ID Gait Challenge Dataset is a collection of video data designed for gait recognition research. It contains videos of 122 subjects, and these subjects can be observed in up to 32 different combinations of variations in various factors.

CMU MoBo: Gross & Shi (2001)

The treadmill in the CMU 3D room was used by 25 people, who were represented in the database. The individuals engaged in four distinct walking patterns: slow walking, fast walking, walking on an incline, and walking while holding a ball. The entire set of subjects was recorded using six high-resolution color cameras evenly positioned around the treadmill.

TUM GAID: Hofmann et al. (2014)

The TUM-GAID dataset comprises data from 305 individuals walking along two distinct indoor paths, and TUM-GAIT stands for TUM Gait from Audio, Image, and Depth. The initial trajectory proceeds from left to right, as the second trajectory progresses from right to left. These recordings were made throughout two sessions, one in January when the subjects were clothed in bulky winter boots and jackets, and the other in April when they were wearing lighter attire. These activities were recorded using a Microsoft Kinect sensor with an image resolution of 640 × 480 pixels and frame rate of 30 frames/s.

CASIA-B: Zheng et al. (2011)

CASIA-B established in January 2005, is an extensive collection of gait information captured from 124 individuals recorded at 25 frames per second with a resolution of 320 × 240 pixels. It included 13,640 video sequences with 11 different viewpoints (ranging from 0° to 180° in 18° increments). This dataset meticulously accounts for three distinct variables: alterations in view angle, clothing, and carrying conditions. In additional, it includes human silhouettes extracted from the corresponding video files.