Temporal Domain Features Research Articles

Improved LSTM-Squeeze Net Architecture for brain activity detection using EEG with improved feature set

Brain activity recognition (BAR) has emerged as most intriguing fields of research in recent days. It poses the ability to transform an extensive number of appliances, together with ICU surveillance, appliance control for the disabled and aged, and the detection of neural diseases. EEG data are the most common representation of brain activity, as they capture the voltage variations of neurons in the brain using non-invasive electrodes implanted on the scalp. Deep learning models play a major role in recognizing brain activity. This work uses EEG inputs to present a new hybrid model-based BAR. Adaptive least mean square (ALMS) is first used to filter the input signal for EEG signals. Next, improved CSP features, morphological features, temporal domain features, and wavelet packet decomposition (WPD) features are extracted. Appropriate characteristics are chosen following feature extraction using Improved Mutual Information (IMI). Then, using an enhanced LSTM (ILSTM)-Squeeze net model, brain actions are detected or recognized. The highest accuracy at LP = 90 is around 0.961. In contrast, the precisions of 0.7385, 0.667, 0.7516, 0.771, 0.686, 0.7124, and 0.7189 were obtained by Dense Net, Mobile Net, Squeeze Net, Res Net, ANN, LSTM, and RNN. The evaluation done proves the efficiency of the proposed model over the other conventional models.

Photoplethysmography Features Correlated with Blood Pressure Changes.

Blood pressure measurement is a key indicator of vascular health and a routine part of medical examinations. Given the ability of photoplethysmography (PPG) signals to provide insights into the microvascular bed and their compatibility with wearable devices, significant research has focused on using PPG signals for blood pressure estimation. This study aimed to identify specific clinical PPG features that vary with different blood pressure levels. Through a literature review of 297 publications, we selected 16 relevant studies and identified key time-dependent PPG features associated with blood pressure prediction. Our analysis highlighted the second derivative of PPG signals, particularly the b/a and d/a ratios, as the most frequently reported and significant predictors of systolic blood pressure. Additionally, features from the velocity and acceleration photoplethysmograms were also notable. In total, 29 features were analyzed, revealing novel temporal domain features that show promise for further research and application in blood pressure estimation.

Assessing expert reliability in determining intracranial EEG channel quality and introducing the automated bad channel detection algorithm

Objective. To evaluate the inter- and intra-rater reliability for the identification of bad channels among neurologists, EEG Technologists, and naïve research personnel, and to compare their performance with the automated bad channel detection (ABCD) algorithm for detecting bad channels. Approach. Six Neurologists, ten EEG Technologists, and six naïve research personnel (22 raters in total) were asked to rate 1440 real intracranial EEG channels as good or bad. Intra- and interrater kappa statistics were calculated for each group. We then compared each group to the ABCD algorithm which uses spectral and temporal domain features to classify channels as good or bad. Main results. Analysis of channel ratings from our participants revealed variable intra-rater reliability within each group, with no significant differences across groups. Inter-rater reliability was moderate among neurologists and EEG Technologists but minimal among naïve participants. Neurologists demonstrated a slightly higher consistency in ratings than EEG Technologists. Both groups occasionally misclassified flat channels as good, and participants generally focused on low-frequency content for their assessments. The ABCD algorithm, in contrast, relied more on high-frequency content. A logistic regression model showed a linear relationship between the algorithm’s ratings and user responses for predominantly good channels, but less so for channels rated as bad. Sensitivity and specificity analyses further highlighted differences in rating patterns among the groups, with neurologists showing higher sensitivity and naïve personnel higher specificity. Significance. Our study reveals the bias in human assessments of intracranial electroencephalography (iEEG) data quality and the tendency of even experienced professionals to overlook certain bad channels, highlighting the need for standardized, unbiased methods. The ABCD algorithm, outperforming human raters, suggests the potential of automated solutions for more reliable iEEG interpretation and seizure characterization, offering a reliable approach free from human biases.

Quantitative Assessment via Multi-Domain Fusion of Muscle Synergy Associated With Upper-Limb Motor Function for Stroke Rehabilitation.

Quantitative assessment of upper limb motor function aids therapists in providing appropriate rehabilitation strategies, which plays an essential role in post-stroke rehabilitation. Traditional assessments, relying on clinical scales or kinematic metrics, often involve subjective scores or are influenced by compensatory strategies. Recently, the use of muscle synergies, representing simplified neuromuscular control, has emerged as a promising approach for post-stroke assessment. In general, muscle synergies are decomposed into two components: synergy vectors and synergy activation. Synergy vectors represent the relative weighting of each muscle within each synergy, that is muscle coordination; synergy activation represents the recruitment of the muscle synergy over time, that is muscle activation strength. Both components are vital for adequately assessing patients' motor function. Therefore, we integrate the spatial domain and temporal domain features extracted from synergy vectors and synergy activation, constructing a multi-domain assessment system using a Random Forest classifier, which may provide great qualitative classification accuracy. Furthermore, a novel functional score is generated from the probabilities belonging to the pathological group. Finally, A study involving ten healthy subjects and ten post-stroke patients validates the proposed method. The experimental results show that the classification accuracy was enhanced to 98.56% by fusing the characteristics derived from different domains, which was higher than that based on spatial domain (94.90%) and temporal domain (91.08%), respectively. Furthermore, the assessment score generated by multi-domain fusion framework exhibited a significant correlation with the clinical score. These promising results show the potential of applying the proposed method to clinical assessments for post-stroke patients.

Multi-scale self-attention approach for analysing motor imagery signals in brain-computer interfaces

BackgroundMotor imagery-based electroencephalogram (EEG) brain-computer interface (BCI) technology has seen tremendous advancements in the past several years. Deep learning has outperformed more traditional approaches, such next-gen neuro-technologies, in terms of productivity. It is still challenging to develop and train an end-to-end network that can sufficiently extract the possible characteristics from EEG data used in motor imaging. Brain-computer interface research is largely reliant on the fundamental problem of accurately classifying EEG data. There are still many challenges in the field of MI classification even after researchers have proposed a variety of methods, such as deep learning and machine learning techniques. MethodologyWe provide a model for four-class categorization of motor imagery EEG signals using attention mechanisms: left hand, right hand, foot, and tongue/rest. The model is built on multi-scale spatiotemporal self-attention networks. To determine the most effective channels, self-attention networks are implemented spatially to assign greater weight to channels associated with motion and lesser weight to channels unrelated to motion. To eliminate noise in the temporal domain, parallel multi-scale Temporal Convolutional Network (TCN) layers are utilized to extract temporal domain features at various scales. ResultOn the IV-2b dataset from the BCI Competition, the suggested model achieved an accuracy of 85.09 %; on the IV-2a and IV-2b datasets from the HGD datasets, it was 96.26 %. Comparison with existing methodsIn single-subject classification, this approach demonstrates superior accuracy when compared to existing methods. ConclusionThe findings suggest that this approach exhibits commendable performance, resilience, and capacity for transfer learning.

A machine learning based depression screening framework using temporal domain features of the electroencephalography signals.

Depression is a serious mental health disorder affecting millions of individuals worldwide. Timely and precise recognition of depression is vital for appropriate mediation and effective treatment. Electroencephalography (EEG) has surfaced as a promising tool for inspecting the neural correlates of depression and therefore, has the potential to contribute to the diagnosis of depression effectively. This study presents an EEG-based mental depressive disorder detection mechanism using a publicly available EEG dataset called Multi-modal Open Dataset for Mental-disorder Analysis (MODMA). This study uses EEG data acquired from 55 participants using 3 electrodes in the resting-state condition. Twelve temporal domain features are extracted from the EEG data by creating a non-overlapping window of 10 seconds, which is presented to a novel feature selection mechanism. The feature selection algorithm selects the optimum chunk of attributes with the highest discriminative power to classify the mental depressive disorders patients and healthy controls. The selected EEG attributes are classified using three different classification algorithms i.e., Best- First (BF) Tree, k-nearest neighbor (KNN), and AdaBoost. The highest classification accuracy of 96.36% is achieved using BF-Tree using a feature vector length of 12. The proposed mental depressive classification scheme outperforms the existing state-of-the-art depression classification schemes in terms of the number of electrodes used for EEG recording, feature vector length, and the achieved classification accuracy. The proposed framework could be used in psychiatric settings, providing valuable support to psychiatrists.

A coarse-to-fine adaptive spatial–temporal graph convolution network with residuals for motor imagery decoding from the same limb

In the field of Brain Computer Interface (BCI) technology, Motor Imagery (MI) plays an important role as a paradigm. One of the primary focuses of this research area lies in exploring the MI of various upper limbs. Decoding MI signals from distinct joints within the same limb poses more intricate challenges in comparison to decoding MI signals originating from different upper limbs. In order to explore more efficient decoding methods, we propose a novel coarse-to-fine classification approach to investigate categorical decoding across three tasks, namely ‘rest’, ‘hand’, and ‘elbow’. This approach consists of two classification stages performed from coarse to fine. In the coarse classification stage, in order to capture the features of both resting state and the moving state in the temporal domain, the EEGNet network with temporal domain convolution is used to extract temporal domain features and classify the original samples into categories of ‘rest’ and ‘move’ (‘hand’, ‘elbow’). In the fine classification stage, the samples of ‘move’ category are segmented by time to form a graph sequence. Then, an adaptive spatial–temporal graph convolutional network with residuals is utilized to extract both temporal and spatial domains’ features from the graph sequence. The proposed algorithm has been validated experimentally on the MI-2 dataset and compared with contemporary methods. Its classification performance is quantified by the average accuracy which achieves a value of 72.21%. Extensive experimental results indicate that the novel coarse-to-fine classification approach is superior to the single classification approach.

A vibration-based 1DCNN-BiLSTM model for structural state recognition of RC beams

The vibration signal analysis-based structural health monitoring (SHM) system has been widely used for bridge condition identification. With the development of artificial intelligence (AI), machine learning (ML) has made numerous breakthroughs in the detection of civil engineering structures. However, conventional data-driven ML methods heavily rely on prior knowledge for their performance. This paper proposes a deep learning (DL) model 1DCNN-BiLSTM for detecting small local structural changes of reinforced concrete (RC) beams, which applies the structure of the Inception module in GoogLeNet to one-dimensional convolution neural networks (1DCNNs) for feature extraction at different scales, and combines the advantages of bidirectional long short-term memory (BiLSTM) modules for processing long time-series data. Firstly, a three-dimensional numerical model of the RC beam was established using finite element software, and the accuracy of the numerical model was verified by comparing it with the test results. Based on the numerical model, tests on RC beams impacted by falling hammers were carried out, and the acceleration signals of the beam in different states were collected as a dataset. The proposed DL model can automatically extract the spatial and temporal domain features in the signals, accurately identify the location of small bridge local variations, the accuracy in the test set is 98.8% compared to 92.6% for a traditional ML model, and has better noise immunity performance and robustness to the missing data than the traditional ML model. Finally, the internal inference process of the model was explored and visualized, illustrating that the model has adaptive learning capabilities.

Research on Mental Workload of Deep-Sea Oceanauts Driving Operation Tasks from EEG Data.

A person's present mental state is closely associated with the frequency and temporal domain features of spontaneous electroencephalogram (EEG) impulses, which directly reflect neurophysiological signals of brain activity. EEG signals are employed in this study to measure the mental workload of drivers while they are operating a vehicle. A technique based on the quantum genetic algorithm (QGA) is suggested for improving the kernel function parameters of the multi-class support vector machine (MSVM). The performance of the algorithm based on the quantum genetic algorithm is found to be superior to that of other ways when other methods and the quantum genetic algorithm are evaluated for the parameter optimization of kernel function via simulation. A multi-classification support vector machine based on the quantum genetic algorithm (QGA-MSVM) is applied to identify the mental workload of oceanauts through the collection and feature extraction of EEG signals during driving simulation operation experiments in a sea basin area, a seamount area, and a hydrothermal area. Even with a limited data set, QGA-MSVM is able to accurately identify the cognitive burden experienced by ocean sailors, with an overall accuracy of 91.8%.

Speech emotion recognition with light gradient boosting decision trees machine

<p>Speech emotion recognition aims to identify the emotion expressed in the speech by analyzing the audio signals. In this work, data augmentation is first performed on the audio samples to increase the number of samples for better model learning. The audio samples are comprehensively encoded as the frequency and temporal domain features. In the classification, a light gradient boosting machine is leveraged. The hyperparameter tuning of the light gradient boosting machine is performed to determine the optimal hyperparameter settings. As the speech emotion recognition datasets are imbalanced, the class weights are regulated to be inversely proportional to the sample distribution where minority classes are assigned higher class weights. The experimental results demonstrate that the proposed method outshines the state-of-the-art methods with 84.91% accuracy on the emo-DB dataset, 67.72% on the Ryerson audio-visual database of emotional speech and song (RAVDESS) dataset, and 62.94% on the interactive emotional dyadic motion capture (IEMOCAP) dataset.</p>

Spatial-Temporal EEG Fusion Based on Neural Network for Major Depressive Disorder Detection.

In view of the major depressive disorder characteristics such as high mortality as well as high recurrence, it is important to explore an objective and effective detection method for major depressive disorder. Considering the advantages complementary of different machine learning algorithms in information mining process, as well as the fusion complementary of different information, in this study, the spatial-temporal electroencephalography fusion framework using neural network is proposed for major depressive disorder detection. Since electroencephalography is a typical time series signal, we introduce recurrent neural network embedded in long short-term memory unit for extract temporal domain features to solve the problem of long-distance information dependence. To reduce the volume conductor effect, the temporal electroencephalography data are mapping into a spatial brain functional network using phase lag index, then the spatial domain features were extracted from brain functional network using 2D convolutional neural networks. Considering the complementarity between different types of features, the spatial-temporal electroencephalography features are fused to achieve data diversity. The experimental results show that spatial-temporal features fusion can improve the detection accuracy of major depressive disorder with a highest of 96.33%. In addition, our research also found that theta, alpha, and full frequency band in brain regions of left frontal, left central, right temporal are closely related to MDD detection, especially theta frequency band in left frontal region. Only using single-dimension EEG data as decision basis, it is difficult to fully explore the valuable information hidden in the data, which affects the overall detection performance of MDD.Meanwhile, different algorithms have their own advantages for different application scenarios. Ideally, different algorithms should use their respective advantages to jointly address complex problems in engineering fields.To this end, we propose a computer-aided MDD detection framework based on spatial-temporal EEG fusion using neural network, as shown in Fig.1. The simplified process is as follows: (1) Raw EEG data acquisition and preprocessing. (2) The time series EEG data of each channel are input as recurrent neural network (RNN), and RNN is used to process and extract temporal domain (TD) features. (3) The BFN among different EEG channels is constructed, and CNN is used to process and extract the spatial domain (SD) features of the BFN. (4) Based on the theory of information complementarity, the spatial-temporal information is fused to realize efficient MDD detection. Fig. 1 MDD detection framework based on spatial-temporal EEG fusion.

Integrating Piecewise Linear Representation and Deep Learning for Trading Signals Forecasting

Trading signal forecasting is an interesting but challenging research topic in the field of financial investment, since the financial market is an nonlinearity and high volatility system influenced by too many factors, and a small improvement in forecasting performance can bring profits. To realize trading signals detection, this paper presents a novel method which integrates piecewise linear representation (PLR) with a deep learning learning framework to predict the financial trading points. Firstly, we utilize PLR to generate a number of turning points (valleys or peaks) from trading data and formulate the trading points prediction as a three-class classification problem. Then, the framework combined a convolutional neural network (CNN) for spatial features extraction and a long short-term memory (LSTM) network for temporal domain features extraction (CNN-LSTM) is used to learn the prediction model between the trading points and the financial time series data. Finally, we conduct a series of experiments among PLR-CNN-LSTM, PLR-CNN-TA and PLR-LSTM on companies of US, Turkey and daily Exchange-Traded Fund (ETFs) to test the performance of our established method. The experiment results show that our proposed method has better model performance and profitability with different investment strategies.

Gated spatial-temporal graph neural network based short-term load forecasting for wide-area multiple buses

Existing short-term bus load forecasting methods mostly use temporal domain features, such as historical loads, to forecast and do not fully consider the influence of unstructured spatial–temporal coupling correlations among multiple bus loads in a wide spatial area on the forecasting results. In this paper, a wide-area multiple bus loads forecasting model is proposed. First, take the rapid computation of the maximal information coefficient (RapidMIC) value between bus loads as edge features, and combine the multiple node feature set to construct a similar-weighted spatial–temporal graph. The constructed similar-weighted spatial–temporal graph is not constrained by geography and grid topology, and can directly reflect the degree of spatial–temporal coupling correlation between bus loads. Second, the neighboring node features of each node in the graph are extracted by the spatial convolution layer (SCL) to achieve full-domain node feature enhancement. Finally, the features extracted were formed into temporal series and input to the gated recurrent unit layer (GRUL) to achieve a wide-area multiple buses short-term load forecast. The research was carried out by selecting the measured load data (15 busbars). The corresponding symmetric mean absolute percentage error (SMAPE) and mean absolute error (MAE) are 3.19 % to 5.89 % and 1.83 megawatts (MW) to 9.8 MW, respectively. Compared with the SMAPE of other methods, the worst evaluation metrics in the test set are improved by 1.82 % to 5.94 %, and it has better robustness in the scene with abnormal load data.

Temporal segment graph convolutional networks for skeleton-based action recognition

Different actions usually emphasize on different parts of a skeleton, even for a specific action, different action stages have the corresponding emphases. Previous studies generally construct the human skeletons as predefined, thus lacking the adaptability to different action modes. In addition, these methods simply employ the padding or truncation operation on the skeleton sequence to fix the sequence length, resulting in additional temporal misalignment problem. In this work, we propose a novel temporal segment graph convolutional networks (TS-GCN) for skeleton-based action recognition. Our model divides the whole sequence into several subsequences. Then GCNs are applied on each subsequence to capture the dynamic information stage by stage, which can align the motion features in temporal domain. Besides, in order to explore the intrinsic features contained in each subsequence, our model introduces a graph-adaptive method to construct an individual graph that can be learned and updated from skeleton data for each subsequence, which increases the generality of graph construction to adapt to different sequences. Extensive experiments are conducted on two standard datasets, NTU-RGB+D and Kinetics. The experimental results demonstrate the effectiveness of the proposed method.

Silicone mask face anti-spoofing detection based on visual saliency and facial motion

Face recognition systems are widely used for target recognition and identity authentication, such as automated teller machines, mobile phones, and entrance guard systems. However, face recognition systems are vulnerable to presentation attacks, such as photo, replay, and 3D mask attacks. In particular, silicone mask attacks pose a greater threat to face recognition systems because high-quality silicone masks do living properties. To promote the development of face anti-spoofing detection algorithms for silicone mask attacks, this paper constructs a Silicone Mask Face Motion Video Dataset (SMFMVD) containing 200 real face videos and 200 silicone mask face videos. These videos include different facial motions collected from 20 subjects. Moreover, inspired by the observation that the silicone mask face’s facial movement is not so natural as the real face, we propose a novel silicone mask face anti-spoofing detection method based on visual saliency and facial motion characteristics. Specifically, we compute the visual saliency map of a given face image by simulating two kinds of eye movement behaviors, namely “gaze” and “saccade”. Then, we propose a saliency-weighted histogram of local binary pattern operator to extract facial texture features in spatial domain and a saliency-guided histogram of oriented optical flow operator to extract facial motion features in temporal domain. Finally, the support vector machine is used to fuse two groups of facial features to distinguish real and spoof faces. Extensive experiments on public and self-built datasets show its superiority over the state-of-the-art methods.

Single-trial detection of EEG error-related potentials using modified power-law transformation

Error processing in the human brain is characterized by a distinct neural response that is elicited when the outcome of an event is not the same as expected. This response is known as error-related potential (ErrP). Detection of error-related potentials from a single-trial data can lead the existing Brain-Computer Interface (BCI) systems one step closer to autonomous system. The non-stationarity of the Electroencephalograph (EEG) signal along with poor spatial resolution and low signal-to-noise ratio make the task of detecting ErrP on a single-trial basis challenging. The objective of this work is to propose a novel ErrP detection method based on temporal domain features of ErrP, particularly, the standard deviation measure. The difference in the standard deviation between the correct and the erroneous trials is verified using a statistical test and this information is subsequently used for ErrP detection. In the proposed method, the EEG electrodes from the frontocentral region of the brain are ranked using standard deviation based ranking criterion and a few top-ranked electrodes will be considered for single-trial detection of ErrP. Then, a modified power-law based transformation is proposed to enhance the discriminability between the correct and error trials. The discriminability is enhanced by transforming the samples lying outside a predefined amplitude range according to power-law equation and the samples within the selected amplitude range are transformed linearly. The modified power-law transformation has two parameters K and γ. The parameter K defines the pre-defined amplitude range and γ is the power in the standard power-law equation. The proposed ErrP detection method is tested on a publicly available EEG database of 10 subjects. The proposed transformation increases the discriminability between the erroneous and correct trials and achieves the best trade-off between the sensitivity and the specificity for detecting ErrP on single-trial basis compared to existing methods. The average sensitivity and specificity achieved is 86.17±9.05% and 92.05±3.69% respectively. The higher sensitivity of detection erroneous trials can be used to faithfully correct the wrongly decoded signal in the EEG based BCI system.

Privacy-preserving IoT Framework for Activity Recognition in Personal Healthcare Monitoring

The increasing popularity of wearable consumer products can play a significant role in the healthcare sector. The recognition of human activities from IoT is an important building block in this context. While the analysis of the generated datastream can have many benefits from a health point of view, it can also lead to privacy threats by exposing highly sensitive information. In this article, we propose a framework that relies on machine learning to efficiently recognise the user activity, useful for personal healthcare monitoring, while limiting the risk of users re-identification from biometric patterns characterizing each individual. To achieve that, we show that features in temporal domain are useful to discriminate user activity while features in frequency domain lead to distinguish the user identity. We then design a novel protection mechanism processing the raw signal on the user’s smartphone to select relevant features for activity recognition and normalise features sensitive to re-identification. These unlinkable features are then transferred to the application server. We extensively evaluate our framework with reference datasets: Results show an accurate activity recognition (87%) while limiting the re-identification rate (33%). This represents a slight decrease of utility (9%) against a large privacy improvement (53%) compared to state-of-the-art baselines.

Learning feature aggregation in temporal domain for re-identification

Person re-identification is a standard and established problem in the computer vision community. In recent years, vehicle re-identification is also getting more attention. In this paper, we focus on both these tasks and propose a method for aggregation of features in temporal domain as it is common to have multiple observations of the same object. The aggregation is based on weighting different elements of the feature vectors by different weights and it is trained in an end-to-end manner by a Siamese network. The experimental results show that our method outperforms other existing methods for feature aggregation in temporal domain on both vehicle and person re-identification tasks. Furthermore, to push research in vehicle re-identification further, we introduce a novel dataset CarsReId74k. The dataset is not limited to frontal/rear viewpoints. It contains 17,681 unique vehicles, 73,976 observed tracks, and 277,236 positive pairs. The dataset was captured by 66 cameras from various angles.

Strain measurement during tensile testing using deep learning-based digital image correlation

This paper describes a novel non-contact strain measurement method defined as deep learning based on digital image correlation (DDIC). In particular, it is very difficult to measure directly displacement of gauge length during tensile testing of thin films. Therefore, we obtained the image data continuously to observe the behavior of the material during tensile testing. The sequential image data obtained at a specific position is assigned to a multi-channel input to train the deep neural network. As a result, the multi-channel image is composed of sequential images obtained along the time domain. Since these images have a correlation with each other along the time domain in each pixel, the neural network learns displacement, including temporal information. The DDIC method originates from a 3D convolutional neural network, which can extract both the spatial and the temporal domain features at the same time. A 3D convolutional filter is used as the feature extraction part of the network in order to effectively learn the input data. The deep learning is an end-to-end learning method and the deformation between two images can be measured regardless of the parameters for the nonlinear deformation of the images. An elastic modulus of 118 GPa, 0.2% yield strength of 941 MPa, ultimate tensile strength of 1108 MPa and fracture strain of 0.02414 are estimated by applying the DDIC method during a tensile test of BeCu thin film. The results of the DDIC method are compared with the displacement sensor data and digital image correlation data.

Visual Recognition of traffic police gestures with convolutional pose machine and handcrafted features

Autonomous vehicles have become a hot spot of the automotive industry, many cities have claimed that autonomous vehicles should be capable of recognizing gestures used by traffic police. Traditional traffic police gesture recognition methods rely on depth-sensor or wearable-devices, which limits their availability in the domain of the intelligent vehicle. Vision-based methods have fewer requirements for distance, but the modeling process is challenging due to the complexity of the visual scenes. Inspired by the recent success in vision-based pose estimation networks such as Convolutional Pose Machine (CPM), in this paper, we propose a novel vision-based human-machine interface to recognize eight kinds of Chinese traffic police gestures and apply it in the real-time recognition tasks. This method integrates a modified CPM network and two kinds of handcrafted features: Relative Bone Length and Angle with Gravity as spatial domain features, and adopt a Long short-term memory (LSTM) network to extract temporal domain features. To train and validate our method, we create a gestures dataset with two hours of traffic police gesture videos, which has 3354 gesture instances. The experiment results show that the proposed method is capable of recognizing traffic police gestures, and is fast enough for online gesture prediction.