Extract Spatiotemporal Features Research Articles

Multifeature extraction based MobileViTv3 model for fish feeding behavior recognition from video streaming

The recognition of fish feeding behavior based on machine vision is essential for optimizing fish feeding strategies and enhancing the efficiency of aquaculture. Building an efficient, multi-feature extraction model for fish feeding recognition, especially on mobile and edge devices, remains a significant challenge. In the paper, we propose a novel multi-feature extraction (MFE)-MobileViTv3 model, which improve MobileViTv3 with the MFE blocks. It can extract spatio-temporal features while obtaining the lightweight characteristic. The MFE block is designed by improving ActionNet, with the frequency channel attention (FCA) and multi-head self-attention (MHSA) mechanisms. It can fully extract spatio-temporal, motion, and channel features from video streaming, thereby further improving the feature extraction capabilities, which subsequently enhances the model's recognition accuracy. The experiments were carried out on an industrial aquaculture farm. We built a dataset of Micropterus salmoides, and then conducted the compare experiments. Compared with C3D, R3D, ResNet50, SlowFast, AlexNet, and MobileNetV3, our model can achieve a classification accuracy of 96.7 % for feeding intensity, with fewer parameters (0.96 M) and FLOPs (8.44G). The results show that the proposed model can effectively recognize fish feeding behavior with fewer parameters. Additionally, we introduce two evaluation metrics for the feeding process: Average Feeding Intensity and Strong Feeding Ratio. The metrics are conducive to the quantitative evaluation of the fish's vigor and health status.

Transformer-based multiview spatiotemporal feature interactive fusion for human action recognition in depth videos

Spatiotemporal feature modeling is the key to human action recognition task. Multiview data is helpful in acquiring numerous clues to improve the robustness and accuracy of feature description. However, multiview action recognition has not been well explored yet. Most existing methods perform action recognition only from a single view, which leads to the limited performance. Depth data is insensitive to illumination and color variations and offers significant advantages by providing reliable 3D geometric information of the human body. In this study, we concentrate on action recognition from depth videos and introduce a transformer-based framework for the interactive fusion of multiview spatiotemporal features, facilitating effective action recognition through deep integration of multiview information. Specifically, the proposed framework consists of intra-view spatiotemporal feature modeling (ISTFM) and cross-view feature interactive fusion (CFIF). Firstly, we project a depth video into three orthogonal views to construct multiview depth dynamic volumes that describe the 3D spatiotemporal evolution of human actions. ISTFM takes multiview depth dynamic volumes as input to extract spatiotemporal features of three views with 3D CNN, then applies self-attention mechanism in transformer to model global context dependency within each view. CFIF subsequently extends self-attention into cross-attention to conduct deep interaction between different views, and further integrates cross-view features together to generate a multiview joint feature representation. Our proposed method is tested on two large-scale RGBD datasets by extensive experiments to demonstrate the remarkable improvement for enhancing the recognition performance.

Early stroke behavior detection based on improved video masked autoencoders for potential patients

Stroke is the prevalent cerebrovascular disease characterized by significant incidence and disability rates. To enhance the early perceive and detection of potential stroke patients, the early stroke behavior detection based on improved Video Masked Autoencoders (VideoMAE) for potential patients (EPBR-PS) is proposed. The proposed method begins with novel time interval-based sampling strategy, capturing video frame sequences enriched with sparse motion features. On the basis of establishing the masking mechanism for adjacent frames and pixel blocks within these sequences, The EPBR-PS employes pipeline mask strategy to extract spatiotemporal features effectively. Then, the local convolution attention mechanism is designed to capture local dynamic feature information, and central to the EPBR-PS is the integration of local convolutional attention mechanism with VideoMAE's multi-head attention mechanism. This integration facilitates the simultaneous leveraging of global high-level semantics and local dynamic feature information. Dual attention mechanism-based method for the fusion of these global and local features is proposed. After that, the optimal parameters of EPBR-PS were determined through the experiment of learning rate and fusion weights of different features. On the NTU-ST dataset, comparative analysis with eight models demonstrated the superiority of EPBR-PS, evidenced by the average recognition accuracy of 89.61%, surpassing that 1.67% over the benchmark VideoMAE. On the HMDB51 dataset, EPBR-PS has Top1 of 71.31%, which is 0.73% higher than that of the VideoMAE, providing the viable behavior detection for perception early signs of potential stroke in the home environment. This code is available at https://github.com/wang-325/EPBR-PS/.

Fall detection method based on spatio-temporal coordinate attention for high-resolution networks

Fall behavior is closely related to the high mortality rate of the elderly, so fall detection has become an important and urgent research area in human behavior recognition. However, the existing fall detection methods, suffer from the loss of detailed action information during feature extraction due to the downsampling operation, resulting in subpar performance when detecting falls with similar behaviors such as lying and sitting. To solve the challenges, this study proposes a high-resolution spatio-temporal feature extraction method based on a spatio-temporal coordinate attention mechanism. The method employs 3D convolutions to extract spatio-temporal features and utilizes gradual down-sampling to generate a multi-resolution sub-network, thus realizing multi-scale fusion and perception enhancement of details. In particular, this study designs a pseudo-3D basic block, which simulates the ability of 3D convolution, to ensure the running speed of the network while controlling the number of parameters. Further, a spatio-temporal coordinate attention mechanism is designed to accurately extract the spatio-temporal positional changes of key skeletal points and the interrelationships among them. Long-term dependencies in horizontal, vertical, temporal directions are captured through three one-dimensional global pooling operations. Then the long-range relationships and channel correlations among features are captured by cascading and slicing operations. Finally, the key information is effectively highlighted by performing dot-multiplication operations between the feature maps from the horizontal, vertical and temporal directions and the input feature maps. Experimental results on three typical public datasets show that the proposed method can better extract motion features and improve the accuracy of fall detection.

HASTF: a hybrid attention spatio-temporal feature fusion network for EEG emotion recognition.

EEG-based emotion recognition has gradually become a new research direction, known as affective Brain-Computer Interface (aBCI), which has huge application potential in human-computer interaction and neuroscience. However, how to extract spatio-temporal fusion features from complex EEG signals and build learning method with high recognition accuracy and strong interpretability is still challenging. In this paper, we propose a hybrid attention spatio-temporal feature fusion network for EEG-based emotion recognition. First, we designed a spatial attention feature extractor capable of merging shallow and deep features to extract spatial information and adaptively select crucial features under different emotional states. Then, the temporal feature extractor based on the multi-head attention mechanism is integrated to perform spatio-temporal feature fusion to achieve emotion recognition. Finally, we visualize the extracted spatial attention features using feature maps, further analyzing key channels corresponding to different emotions and subjects. Our method outperforms the current state-of-the-art methods on two public datasets, SEED and DEAP. The recognition accuracy are 99.12% ± 1.25% (SEED), 98.93% ± 1.45% (DEAP-arousal), and 98.57% ± 2.60% (DEAP-valence). We also conduct ablation experiments, using statistical methods to analyze the impact of each module on the final result. The spatial attention features reveal that emotion-related neural patterns indeed exist, which is consistent with conclusions in the field of neurology. The experimental results show that our method can effectively extract and fuse spatial and temporal information. It has excellent recognition performance, and also possesses strong robustness, performing stably across different datasets and experimental environments for emotion recognition.

An Ultra‐Short‐Term Wind Power Prediction Method Based on Spatiotemporal Characteristics Fusion

ABSTRACTAiming at the problem that the existing ultra‐short‐term wind power prediction methods lack consideration of the spatial correlation characteristics of wind farms, resulting in insufficient prediction accuracy, an ultra‐short‐term wind power prediction method based on spatiotemporal characteristics fusion is proposed in this article. First, the fluctuation difference of the time window of wind power is input into the K‐means clustering algorithm to cluster wind farms into several clusters based on power fluctuation similarity. Then, the principal component analysis algorithm is used to reduce the dimensionality of numerical weather prediction data combinations in different regions to reduce the impact of redundant information on modeling accuracy. Finally, a convolutional long‐short‐term memory neural network is designed to extract spatiotemporal features of wind power data and output prediction results. The experimental verification on 18 wind farms in a province in China shows that the proposed wind power prediction method has an average root mean square error of only 0.1257 and has certain applicability.

Identification of plagioclase extinction-angle features from polarized images using deep neural network.

Plagioclase is a principal component of the Earth's crust, whose compositional and structural analysis is vital for understanding the crust's construction and evolution. Accurate identification of extinction angle features plays an important role in determining the sodium-calcium content in plagioclase. Manual evaluation of these extinction angle features is tedious and dependent on human expertise. Additionally, current image recognition methods for identifying plagioclase extinction angles face challenges such as information loss, weak stripe features, and difficulty in capturing long-range spatiotemporal features. To address these challenges, we propose an extinction angle identification neural network called AFI-Net, which utilizes polarized image sequences for the accurate detection of plagioclase's extinction angle features. AFI-Net combines a 2D convolutional neural network with a Transformer. Initially, a stripe attention module is developed to enhance the network's ability to detect stripe features. Building upon this module, a 2D backbone network is designed to efficiently extract spatial features from polarized images. The spatial features are then fed into a customized Transformer-based module to extract spatiotemporal features. Ultimately, these spatiotemporal features are used to accurately identify the extinction angle features of plagioclase. Extensive quantitative and qualitative experimental results demonstrate that AFI-Net achieves high accuracy and stability in recognizing the extinction angle features of plagioclase, showing significant superiority over current advanced recognition methods.

Gear Classification in Skating Cross-Country Skiing Using Inertial Sensors and Deep Learning.

The aim of this current work is to identify three different gears of cross-country skiing utilizing embedded inertial measurement units and a suitable deep learning model. The cross-country style studied was the skating style during the uphill, which involved three different gears: symmetric gear pushing with poles on both sides (G3) and two asymmetric gears pushing with poles on the right side (G2R) or to the left side (G2L). To monitor the technique, inertial measurement units (IMUs) were affixed to the skis, recording acceleration and Euler angle data during the uphill tests performed by two experienced skiers using the gears under study. The initiation and termination points of the tests were controlled via Bluetooth by a smartphone using a custom application developed with Android Studio. Data were collected on the smartphone and stored on the SD memory cards included in each IMU. Convolutional neural networks combined with long short-term memory were utilized to classify and extract spatio-temporal features. The performance of the model in cross-user evaluations demonstrated an overall accuracy of 90%, and it achieved an accuracy of 98% in the cross-scene evaluations for individual users. These results indicate a promising performance of the developed system in distinguishing between different ski gears within skating styles, providing a valuable tool to enhance ski training and analysis.

Emotion Recognition Using EEG Signals and Audiovisual Features with Contrastive Learning.

Multimodal emotion recognition has emerged as a promising approach to capture the complex nature of human emotions by integrating information from various sources such as physiological signals, visual behavioral cues, and audio-visual content. However, current methods often struggle with effectively processing redundant or conflicting information across modalities and may overlook implicit inter-modal correlations. To address these challenges, this paper presents a novel multimodal emotion recognition framework which integrates audio-visual features with viewers' EEG data to enhance emotion classification accuracy. The proposed approach employs modality-specific encoders to extract spatiotemporal features, which are then aligned through contrastive learning to capture inter-modal relationships. Additionally, cross-modal attention mechanisms are incorporated for effective feature fusion across modalities. The framework, comprising pre-training, fine-tuning, and testing phases, is evaluated on multiple datasets of emotional responses. The experimental results demonstrate that the proposed multimodal approach, which combines audio-visual features with EEG data, is highly effective in recognizing emotions, highlighting its potential for advancing emotion recognition systems.

AONet: Attention network with optional activation for unsupervised video anomaly detection

AbstractAnomaly detection in video surveillance is crucial but challenging due to the rarity of irregular events and ambiguity of defining anomalies. We propose a method called AONet that utilizes a spatiotemporal module to extract spatiotemporal features efficiently, as well as a residual autoencoder equipped with an attention network for effective future frame prediction in video anomaly detection. AONet utilizes a novel activation function called OptAF that combines the strengths of the ReLU, leaky ReLU, and sigmoid functions. Furthermore, the proposed method employs a combination of robust loss functions to address various aspects of prediction errors and enhance training effectiveness. The performance of the proposed method is evaluated on three widely used benchmark datasets. The results indicate that the proposed method outperforms existing state‐of‐the‐art methods and demonstrates comparable performance, achieving area under the curve values of 97.0%, 86.9%, and 73.8% on the UCSD Ped2, CUHK Avenue, and ShanghaiTech Campus datasets, respectively. Additionally, the high speed of the proposed method enables its application to real‐time tasks.

AMTrack:Transformer tracking via action information and mix-frequency features

Nowadays, Transformer-based visual tracking algorithms have been developing quickly because of the self-attention mechanism of Transformer, which has the capability to model global information. Although the self-attention mechanism in Transformer can effectively capture long-range dependencies in feature space, they only use flattened two-dimensional features and are unable to capture long-range temporal dependencies. Furthermore, since the self-attention in Transformer functions as a low-pass filter, it picks up on low-frequency features of the target while ignoring high-frequency features. This research suggests a Transformer tracker based on action information and mix-frequency features (AMTrack) to address these problems. Specifically, to address the lack of temporal remote dependencies, we introduce the target action aware module and the target action offset module. The target action aware module sets up several pathways to extract spatio-temporal, channel, and motion feature independently. In contrast, the target action offset module computes the target’s offset information by computing relative feature maps. Furthermore, in order to address the imbalance between high and low frequency features, we propose a mix-frequency attention and multi-frequency self-attention convolutional block. The mix-frequency attention uses high-frequency features within partitioned local windows as input for the high-frequency branch and average-pooled low-frequency features as the input for the low-frequency branch, calculating attention scores in both branches respectively. The multi-frequency self-attention convolutional block uses self-attention to capture low-frequency features and convolution to capture high-frequency features. Extensive experiments are carried out on eight challenging tracking datasets (e.g., OTB100 (Object Tracking Benchmark 100), NFS (Need For Speed), UAV123 (Unmanned Aerial Vehicles 123), TC128 (Temple Color 128), VOT2018 (Visual Object Tracking 2018), LaSOT (Large-scale Single Object Tracking), TrackingNet (Tracking Network), GOT-10k (Generic Object Tracking-10k)), and the experimental results show that our tracker achieves excellent tracking performance when compared with several state-of-the-art tracking algorithms. The experimental results show that on LaSOT, the success rates AUC (Area Under Curve), PNorm, and P reach 65.8%, 69.2%, and 68.0%, respectively, where the AUC value is 2.1% higher than the baseline algorithm TrDiMP (Transformer Discriminative Model Prediction). On other datasets, our tracker also achieves excellent tracking performance.

Trajectory Privacy-Protection Mechanism Based on Multidimensional Spatial–Temporal Prediction

The popularity of global GPS location services and location-enabled personal terminal applications has contributed to the rapid growth of location-based social networks. Users can access social networks at anytime and anywhere to obtain services in the relevant location. While accessing services is convenient, there is a potential risk of leaking users’ private information. In data processing, the discovery of issues and the generation of optimal solutions constitute a symmetrical process. Therefore, this paper proposes a symmetry–trajectory differential privacy-protection mechanism based on multi-dimensional prediction (TPPM-MP). Firstly, the temporal attention mechanism is designed to extract spatiotemporal features of trajectories from different spatiotemporal dimensions and perform trajectory-sensitive prediction. Secondly, class-prevalence-based weights are assigned to sensitive regions. Finally, the privacy budget is assigned based on the sensitive weights, and noise conforming to localized differential privacy is added. Validated on real datasets, the proposed method in this paper enhanced usability by 22% and 37% on the same dataset compared with other methods mentioned, while providing equivalent privacy protection.

Optimized deep learning modelling for predicting the diffusion range and state change of filling projects

Concealment of filling constructions poses significant challenges for quality assurance in filling engineering. Direct surveillance of fill dispersal currently remains infeasible, while conventional detection techniques suffer deficiencies in efficiency. This research proposes a framework integrating elastic wave monitoring and hybrid deep learning for predictive modelling of filling state transitions and diffusion range. During the sand filling of the immersed tunnel, elastic wave data is collected via elastic wave testing, and the response energy characteristic is derived through time-domain analysis. The trends in elastic wave response energy are correlated with three filling states: free diffusion, accumulation, and filled state, using Seasonal and Trend decomposition using Loess (STL) for seasonal trend analysis. Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) are utilized to extract spatiotemporal features from the response energy trends, facilitating accurate prediction of the trends’ development and the sand filling state over time. The performances of the proposed strategy are illustrated through an application to the case study of the sand filling construction of the Chebeilu immersed tunnel. The CNN + LSTM model with the proposed strategy gave excellent results (MAE 0.0663, MSE 0.0071, RMSE 0.0845). The model can predict fill state changes and quantify diffusion radii to optimize and guide the construction process.

Research on Surgical Gesture Recognition in Open Surgery Based on Fusion of R3D and Multi-Head Attention Mechanism

Surgical gesture recognition is an important research direction in the field of computer-assisted intervention. Currently, research on surgical gesture recognition primarily focuses on robotic surgery, with a lack of studies in traditional surgery, particularly open surgery. Therefore, this study established a dataset simulating open surgery for research on surgical gesture recognition in the field of open surgery. With the assistance of professional surgeons, we defined a vocabulary of 10 surgical gestures based on suturing tasks in open procedures. In addition, this paper proposes a surgical gesture recognition method that integrates the R3D network with a multi-head attention mechanism (R3D-MHA). This method uses the R3D network to extract spatiotemporal features and combines it with the multi-head attention mechanism for relational learning of these features. The effectiveness of the R3D-MHA method in the field of open surgery gesture recognition was validated through two experiments: offline recognition and online recognition. The accuracy at the gesture instance level for offline recognition was 92.3%, and the frame accuracy for online recognition was 73.4%. Finally, its performance was further validated on the publicly available JIGSAWS dataset. Compared to other online recognition methods, the accuracy improved without using additional data. This work lays the foundation for research on surgical gesture recognition in open surgery and has significant applications in process monitoring, surgeon skill assessment and educational training for open surgeries.

An evolutionary deep learning model based on XGBoost feature selection and Gaussian data augmentation for AQI prediction

Accurate prediction of air quality is crucial for ensuring the scientific validity and effectiveness of air pollution control measures. This study proposes a combined deep learning (DL) model (XGBoost-GDA-TCN-IMRFO-GRU) for predicting hourly air quality index (AQI) data in four cities. The model integrates Extreme gradient boosting (XGBoost) for feature selection, Gaussian data augmentation (GDA), improved manta ray foraging optimization (IMRFO) algorithm, temporal convolutional network (TCN), and gated recurrent unit (GRU). XGBoost calculates the scores of pollutant gases affecting AQI, selecting the top four important pollutants (PM2.5, PM10, NO2, O3) based on their importance rankings. GDA enhances the robustness of the DL models and addresses the limitations of insufficient and overfitting training datasets. Additionally, the IMRFO algorithm, with two improved strategies, is applied to enhance the GRU model. TCN extracts spatiotemporal features of AQI, while GRU constructs a temporal model for efficient computations. Compared to eleven benchmark models, the proposed model demonstrates superior performance in terms of MAE, RMSE, MAPE, and NSE, achieving high accuracy and optimal prediction performance. Specifically, the XGBoost-GDA-TCN-IMRFO-GRU model reduces RMSE, MAE, and MAPE by 33–60 %, 39–68 %, and 39–66 %, respectively, compared to the TCN model. Therefore, the XGBoost-GDA-TCN-IMRFO-GRU model can provide reliable early warnings for air quality, which is of great significance for air pollution prevention and the sustainable development of society.

PI-STGnet: Physics-integrated spatiotemporal graph neural network with fundamental diagram learner for highway traffic flow prediction

At present, traffic state prediction primarily relies on purely data-driven methods, ignoring the incorporation of physical constraints within the field of traffic flow. Taking this as a starting point, this paper endeavors to embed the physical mechanism of the traffic flow fundamental graph into the deep learning model, and proposes a physics-integrated spatiotemporal graph neural network with fundamental diagram learner (PI-STGnet) for highway traffic flow prediction. The PI-STGnet mainly comprises a semantic enhancement module (SE), a spatiotemporal correlation extraction module (ST-Block), and a fundamental diagram learner module (FD-learner). These modules are strategically designed to sequentially enhance the contextual semantic relationships of the traffic flow and speed inputs, extract intricate spatiotemporal correlations of traffic states, and adaptively learn the dynamic evolution of the physical relationships between traffic flow and speed. The real-world dataset employed to assess the performance of the PI-STGnet is sourced from the monitoring data of highway gantry sensors located in Ningde City, Fujian Province, China. The experimental results demonstrate that the proposed PI-STGnet has the advantages to extract spatiotemporal features and achieve prediction, surpassing the accuracy of cutting-edge baseline models. In summary, as an exploratory work, this paper provides a novel approach that driven by the fusion of deep learning and physical mechanisms for traffic flow prediction to a certain extent.

PulseNet: Multi-task learning-based non-contact pulse condition diagnosis using multi-scale fusion and transformer

The pulse condition, also called mai xiang, a significant diagnostic indicator in traditional Chinese medicine (TCM), has increased demand for intelligent diagnosis in recent years. However, the existing intelligent pulse condition diagnosis is mostly sensor-based or contact-dependent, resulting in numerous inconveniences, especially in mobile healthcare. Recently, remote photoplethysmography(rPPG) with the information captured from facial videos has been widely used in non-contact physiological measurements, such as blood pressure and heart rate monitoring. Inspired by these studies, a novel end-to-end rPPG-based method, named PulseNet, is proposed for non-contact pulse condition diagnosis in this paper. In PulseNet, the transformer is employed to extract spatio-temporal features from facial videos, multi-scale fusion is adopted to enhance the feature representation progressively, and multi-task learning is proposed for final pulse condition diagnosis. Besides, we propose the spatio-temporal difference attention (STDA) block in the transformer to aggregate spatio-temporal difference clues of the local region to provide fine-grained spatio-temporal context. We conduct extensive experiments with the proposed method. The results show that PulseNet outperforms the state-of-the-art rPPG-based methods. To the best of our knowledge, our method is the first attempt to detect pulse conditions with the information captured from facial videos, making it a promising non-contact diagnosis approach for mobile healthcare in TCM.

Single-Trial Detection and Classification of Event-Related Optical Signals for a Brain-Computer Interface Application.

Event-related optical signals (EROS) measure fast modulations in the brain's optical properties related to neuronal activity. EROS offer a high spatial and temporal resolution and can be used for brain-computer interface (BCI) applications. However, the ability to classify single-trial EROS remains unexplored. This study evaluates the performance of neural network methods for single-trial classification of motor response-related EROS. EROS activity was obtained from a high-density recording montage covering the motor cortex during a two-choice reaction time task involving responses with the left or right hand. This study utilized a convolutional neural network (CNN) approach to extract spatiotemporal features from EROS data and perform classification of left and right motor responses. Subject-specific classifiers trained on EROS phase data outperformed those trained on intensity data, reaching an average single-trial classification accuracy of around 63%. Removing low-frequency noise from intensity data is critical for achieving discriminative classification results with this measure. Our results indicate that deep learning with high-spatial-resolution signals, such as EROS, can be successfully applied to single-trial classifications.

Research on investment project risk prediction and management based on machine learning

In the era of digital economy, digital transformation is an inevitable choice in line with current economic development and national policy trends. Enterprises use new generation digital technologies such as big data, blockchain, cloud computing, artificial intelligence, and financial technology to apply these technologies to various production and business activities. The research on project risk management has gradually introduced the method of intelligent decision-making, using technologies such as big data and artificial intelligence to identify and analyze risks. We use learning models and ensemble learning techniques to predict and manage the risks of investment projects. Through long short-term memory networks (LSTMs), we are able to effectively extract spatiotemporal features from historical investment data to predict future risk dynamics. In order to further improve the accuracy of prediction and the robustness of the model, we introduced the Gradient Booster (GBM) ensemble learning method. These integrated technologies not only optimize the overall performance of the model, but also enhance the adaptability and forecast accuracy of various market changes. In the experimental analysis, we compare the performance of multiple models on real-world investment datasets, and the results show that the ensemble learning method has significant advantages in risk prediction accuracy.

Predicting spatio-temporal traffic flow: a comprehensive end-to-end approach from surveillance cameras

Traffic flow forecasting is an essential aspect of intelligent traffic management. It enables timely and proactive management of modern transport systems, increasing efficiency and resilience. However, accurately predicting short-term traffic flow is challenging due to its uncertain and interconnected nature. Traditional methods like loop detectors and high-resolution cameras have limited scalability. To address this, we propose a two-stage approach using low-resolution surveillance cameras. The first stage involves a vision-based data extraction module with calibration, vehicle detection, and tracking. Integration of Region of Interest, fine-tuning, and post-processing improves the robustness of low-resolution videos. In the second stage, a novel deep learning model extracts spatio-temporal features from historical traffic data for short-term flow prediction. The proposed model outperforms the STGCN model, achieving an 11.19% increase in MAE, a 12.37% improvement in RMSE and a 4.97% reduction in inference time. These advances highlight its potential for further research and applications in the field.