International Journal of Intelligent Systems Technologies and Applications (9 papers in press)

Regular Issues

• Integrating Incremental and Continuous Learning into an Online Emotion Recognition Model  
  by Caihua Chen, Long Xinyuan 
  Abstract: In order to overcome the problems of low accuracy, F1 score, and long recognition time in traditional online emotion recognition methods, a new online emotion recognition model design method integrating incremental and continuous learning is proposed. Systematically collect and preprocess text, speech, and image data. The incremental learning model adopts a dynamic preservation set strategy based on scaling and translation selection method to select informative samples from multimodal data. Build an online emotion recognition model based on continuous learning using the selected samples. This model introduces a positive definite symmetric matrix to constrain the gradient update direction, and achieves feature decoupling and reuse through a task identifier driven feature encoding layer to obtain online emotion recognition results. The experimental results show that the accuracy of the proposed method can reach up to 98%, with an F1 value maintained above 0.95 and an average recognition time of only 1.24 s.
  Keywords: Incremental learning; Continuous learning; Online emotions Recognition model; Dynamic preservation set; Positive definite symmetric matrix; feature encoding.
  DOI: 10.1504/IJISTA.2026.10077230
   
  
• Research on a Computer Vision-Based Quantifying System for Movement Standardisation in Baduanjin Exercises for the Elderly  
  by Xinyu Wang, Tao Wu, Yanzhao Guo, Junxiang Qiu, Mingli Zhou 
  Abstract: This study develops a computer vision system to quantify movement standardisation in elderly Baduanjin exercise .Using smartphone cameras and OpenPose framework with human pose estimation (HPE), the system detects body keypoints through deep convolutional neural networks. It calculates standardized scores (0-1 scale) via weighted Euclidean distances, specifically adapted for elderly physical traits like mild hunching. Error reduction techniques (outlier filtering/interpolation) improved detection accuracy by 15%. In a 12-week trial with 52 participants, the system tracked significant movement standardisation improvements from 0.871 to 0.876 (P<0.0167). Compared to manual assessments, this solution eliminates subjective bias while enabling home-based exercise monitoring through accessible technology. The system fills a technical gap in geriatric rehabilitation by providing precise data support for personalized training plans. Future development may incorporate 3D pose estimation to enhance clinical utility.
  Keywords: Computer Vision; Human Pose Estimation (HPE); Movement Standardisation Quantification.
  DOI: 10.1504/IJISTA.2026.10077781
   
  
• Research on constant temperature control strategy for nonlinear large time-delay fluid systems based on fuzzy Smith predictor  
  by Panpan Li, Chen Niu 
  Abstract: Precise temperature control of fluid systems is critical in industrial and agricultural applications but is often challenged by large thermal inertia, pure transport delays, and nonlinear parameter variations. Conventional PID controllers and standard Smith predictors often fail to provide satisfactory performance under such complex conditions, suffering from overshoot or instability due to model mismatch. To address these issues, this paper proposes a novel constant temperature control strategy based on a Fuzzy-Smith predictor. The proposed method integrates a fuzzy logic inference engine to adaptively tune the PID parameters online, compensating for system nonlinearities, while a modified Smith structure is employed to effectively mitigate the impact of time delays. Extensive numerical simulations and physical experiments on an STM32-based fluid heating platform were conducted. The results demonstrate that the proposed strategy achieves a settling time reduced by approximately 40% compared to conventional PID, with negligible overshoot (<0.5 C) and superior robustness against parameter perturbations (30% mismatch) and external disturbances.
  Keywords: fluid temperature control; large time delay; nonlinear system; fuzzy logic control; Smith predictor; adaptive control.
  DOI: 10.1504/IJISTA.2026.10078186
   
  
• Study on Multimodal Behaviour Detection of Students in Chinese Classrooms Based on the Internet of Things  
  by Shuai Yu, Linlin Zhu 
  Abstract: Detecting multimodal behaviors of students in Chinese classrooms is of great significance for comprehensively and accurately grasping students' learning status and optimizing teaching strategies. A multimodal behavior detection method of students in Chinese classrooms based on the Internet of things is proposed. Firstly, with the support of IoT technology, cameras and pyroelectric sensors are used to collect data,which is then fused and processed to obtain high-quality data. Secondly, the collected images undergo denoising, grayscale transformation, enhancement, and other operations. Finally, the processed data is input into an improved YOLOv5 model, where the Backbone network extracts features and the Neck layer fuses low-level and high-level information to generate three-scale feature maps.Each grid predicts bounding boxes and outputs relevant information,achieving multimodal behavior detection of students in Chinese classrooms.Experimental results show that the average detection accuracy of this method reaches 99.65%, with an AUC value consistently higher than 0.95
  Keywords: Internet of Things; Chinese classrooms; Student; Multimodal behavior; detection.
  DOI: 10.1504/IJISTA.2026.10078466
   
  
• HESRL: Enhancing Social Robot Navigation Safety Through Hazard-Aware Exploration  
  by Yingao Fu, Hao Fu 
  Abstract: Reinforcement learning-based social navigation is promising for human-robot coexistence. However, the trial-and-error nature of reinforcement learning (RL) often leads to the emergence of unsafe behaviours during training. To address this, we propose a hazard-exploration safe RL (HESRL) by integrating safe exploration filter (SEF) with hazard exploration virtual robot (HEVR). The SEF integrates dynamic control barrier function with model predictive control to adjust the RL-generated actions, ensuring adherence to safety constraints. Additionally, we introduce a HEVR accompanied by a risk-driven reward function to allow the robot to acquire experience within hazardous areas. Simulation results demonstrate that our algorithm outperforms state-of-the-art methods in terms of exploration safety and ensuring post-convergence model performance. Specifically, our algorithm achieves 0% collision rate and 1% discomfort frequency, and improves the success rate by 1.227.2% over baseline methods. Furthermore, the experimental results on a TurtleBot4 show that our algorithm enables the robot to navigate to its goal.
  Keywords: social robot navigation; safe reinforcement learning; hazard-aware exploration; risk-driven reward.
  DOI: 10.1504/IJISTA.2026.10079136
   
  
• Spatio-Temporal Hierarchical Network for Human Pose Recognition Using Millimetre-Wave Radar  
  by Gang Yang, Ya Gao, Jintao Shi 
  Abstract: To enhance the accuracy and real-time performance of human pose recognition using millimeter-wave radar, an end-to-end spatio-temporal hierarchical feature fusion network (STHNet) is proposed. Multi-scale feature extraction is conducted through cascaded spatial and temporal convolutions, where a temporal convolution module as an independent unit links spatial representations with long-term dependencies, improving extraction of short-term motion cues and long-range semantics from sparse radar sequences. An adaptive feature selection mechanism is introduced, where learnable weights dynamically emphasize key spatio-temporal features, improving the discrimination of subtle motion variations. The end-to-end training strategy removes handcrafted feature design and enables a direct mapping from raw radar signals to pose categories. Experimental results show that STHNet achieves 99.5% overall recognition accuracy across six daily postures and a 100% recall rate for fall detection. Moreover, the average inference latency is 25.61 ms, with 95% of samples processed within 35.433 ms, demonstrating satisfactory real-time performance for practical deployment.
  Keywords: millimeter-wave radar; human pose recognition; end to end; deep learning; spatio-temporal features.
  DOI: 10.1504/IJISTA.2026.10079881
   
  
• Explainable AI-Driven Face Morphing and Demorphing Using Face Recognition Techniques Based on FSGAN  
  by M.K.Mohamed Faizal, S. Geetha, A. Barveen 
  Abstract: Face morphing attacks are a threat to security systems that use faces to identify people. This is because people can combine faces from individuals to gain unauthorised access. This study proposes a method for face morphing and demorphing using face swapping generative adversarial network (FSGAN). Facial images are encoded using a pre-trained ArcFace model. These images are created by combining features from two identities is done at a 70:30 ratio. The result is a realistic hybrid face. It was tested on the CelebA dataset. The results are as good in accuracy of morphing, morph detection and identity recovery rate. 3.12% is considered to be the false acceptance rate of this system and also considered to have a better performance than other technologies like StarGAN and DeepFakeGAN. As a whole, this proposed system increases the transparency and helps in automated biometric verification system that paves way for morphed image generation, identity recovery and identity detection by showing the improved potential in real-time applications by providing greater security.
  Keywords: Biometric Security; Face Recognition; Face Morphing; Morph Detection; Demorphing; ArcFace; FSGAN; SHAP; Grad-CAM XGBoost; Explainable AI.
  DOI: 10.1504/IJISTA.2026.10079886
   
  
• Quantum Inspired Siamese Convolutional Forward Harmonic Net with THEuCaLoss function for Percentage of Nutrient Deficiency Estimation and Fertilizer Recommendation using cotton plants leaf images  
  by Swapnil S. Ayane, Mukesh Tiwari 
  Abstract: In this work, we introduce a novel model, the Tanimoto Harmonic Euclidean Cauchy Loss Net (THEuCaLossNet), for accurate identification of nutrient deficiencies in cotton plants. The process begins with image pre-processing using a Non-Local Means (NLM) filter, followed by segmentation through region-based Online Selective Examination (ROSE). Image augmentation techniques such as cropping, resizing, and Grid Mask enhance model performance. Feature extraction is conducted using Locally Adaptive Regression Kernel (LARK) and EfficientNet. These features are passed into the Quantum Siamese Convolutional Forward Harmonic Network (QuSCFHNet), combining Quantum-Inspired Convolutional Neural Networks (CNNs) and Siamese CNNs. The network is trained using Cauchy-Schwarz Divergence, Euclidean loss, and harmonic functions. Based on the results, organic fertilizer recommendations are provided. The model achieves 92.66% accuracy, 92.74% True Positive Rate (TPR), and 92.43% True Negative Rate (TNR), proving its effectiveness.
  Keywords: Non-Local Means; Region-based Online Selective Examination; Locally Adaptive Regression Kernel; Quantum-Inspired Convolutional Neural Networks; Siamese Convolutional Neural Networks.
  DOI: 10.1504/IJISTA.2026.10080503
   
  
• Implementation of a Voice Pathology Detection and Severity Analysis System for Parkinsons Disease using Convolutional Neural Networks  
  by Anitha Sankaran, Shaik Fardeen Hussain, Vinoth B, Dhanvantraj M, Lakshmi Sutha Kumar 
  Abstract: A simple yet effective CNN model is proposed in this paper for Parkinsons disease detection. The proposed Convolutional Neural Network (CNN) is fed with Mel Frequency Cepstral Coefficients (MFCC) for detecting Parkinsons disease from speech signals and gives an accuracy of 86.67%. The DisVoice framework is used to extract the phonation and prosodic characteristics from speech signals. The MFCCs, which are an articulation parameter, along with the phonation and prosody features, provide a high-dimensional feature vector comprising of 171 features. The SelectKBest feature selection method is employed to optimize the model performance and identify the most significant features for predicting the severity of Parkinsons disease. The selected significant features are fed to the three different machine learning algorithms for determining the severity analysis of Parkinsons disease, of which the XGBoost algorithm performs the best with a recognition accuracy of 83.33%
  Keywords: Parkinson’s Disease; PC GITA; Mel Frequency Cepstral Coeffecients; CNN; XGBoost.