Md Baharul Islam

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition that affects about 1 in 100 children worldwide. Early detection supports timely intervention, but many families still face long waits and uneven access to specialists. In low-resource settings, diagnostic delays can reach around 24 months. In this paper, we synthesize technology-enabled approaches for ASD detection across four data modalities. These include facial expression and gaze analysis, voice and audio assessment, sensor and brain-imaging signals, and multimodal integration. We summarize the main methods and report performance ranges from 68.5% to 98.9% accuracy across studies. We find that deep learning approaches often perform best, especially when they use transfer learning and multimodal fusion. Compared with prior reviews that focus on a single modality or emphasize algorithms alone, we connect performance evidence to clinical readiness factors such as cost, accessibility, scalability, validation, and real-world deployment constraints. At the same time, key barriers remain. Common issues include limited dataset diversity, weak cross-cultural generalizability, a lack of clinical validation, and ethical risks in real-world deployment. We close by outlining practical directions for building detection tools that are clinically useful and globally accessible.

Corporate sustainability reports are often difficult for non-expert audiences to understand because they are typically presented as static, number-dense PDF documents. This paper presents an interactive Augmented Reality (AR) system that transforms Herc Rentals' Corporate Citizenship Report data into spatially anchored 3D visualizations on a commodity Android smartphone. Developed using Unity and AR Foundation, the system detects the corporate logo and dynamically renders three scaled emission spheres representing Scope 1, Scope 2, and Scope 3 greenhouse gas emissions, totaling 358,841 MT CO 2 e in 2024. By tapping a sphere, users can access contextual panels that provide category-level breakdowns and carbon offset equivalencies derived from U.S. Environmental Protection Agency (EPA) greenhouse gas equivalency factors. A preliminary user evaluation with six participants (N = 6), using the System Usability Scale (SUS) and NASA Task Load Index (NASA-TLX), produced a mean SUS score of 77.9, surpassing the target threshold of 75, along with generally low workload ratings across NASA-TLX dimensions. These findings suggest that AR-based sustainability visualization can improve comprehension, engagement, and environmental awareness in corporate reporting contexts.

This paper presents a feasibility analysis of saliency detection algorithms for stereo video retargeting in assistive augmented reality (AR) systems. Targeting the Magic Leap 2 headset, we benchmark a diverse set of classical and lightweight deep learning models, including AIM, SUN, IKN, BMS, U 2 -Net, and SAM-Net, evaluating their performance in terms of segmentation accuracy, inference speed, and deployment efficiency. Unlike prior approaches that rely on computationally intensive deep neural networks, our goal is to identify models compatible with the limited hardware acceleration available in wearable AR devices. A personalized visual mask, calibrated through user-specific visual profiles, guides the adaptive relocation of salient regions into areas of preserved vision, enabling dynamic field-of-view compensation for individuals with peripheral vision impairments. Experimental results demonstrate that U 2 -Net Lite and SAM-Net offer a practical balance between accuracy and computational cost, while classical methods like BMS remain viable for real-time AR integration. This work contributes to the development of low-latency, vision-aware AR systems, advancing stereo-retargeting pipelines for accessibility-focused applications. All benchmarking code, evaluation scripts, and model wrappers are available upon request.

Cybersickness remains a major barrier to comfortable virtual reality (VR) use, particularly during rapid changes in motion. We present a jerk-triggered bone-conducted vibration (BCV) system that delivers brief BCV stimuli shortly after the peak of linear jerk, targeting the transition where linear acceleration begins to decrease. This timing is designed to induce an inverse coupling effect by introducing BCV as acceleration decays. We develop a low-cost hardware platform and a software pipeline that enables real-time triggering and seamless compatibility with existing SteamVR applications. To support practical deployment, we introduce a lightweight earhook mounting concept that clips directly onto a VR headset and maintains consistent contact with the mastoid without headbands, calibration, or user-specific fitting. We report engineering validation of system operation, including end-to-end triggering behavior and integration reliability, and outline a future controlled human-subject study to evaluate effectiveness for cybersickness mitigation.

Thermal surveillance systems deployed over months must remain accurate despite environmental thermal drift induced by seasonal weather variation, ambient temperature changes, and sensor recalibration. We present results on the Long-Term Thermal Detection (LTDv2) benchmark using a YOLOv11m detector adapted for single-channel long-wave infrared (LWIR) imagery via a thermally consistent training pipeline: we disable RGB-centric color augmentations (e.g., HSV hue/saturation) that are invalid for grayscale LWIR and emphasize intensity-domain perturbations that better reflect long-term drift, retaining controlled brightness variation to emulate sensor aging and scene temperature shifts while keeping the remainder of the YOLOv11 training recipe unchanged for stable optimization. On the 8-month LTDv2 benchmark, our approach achieves a weighted score of 0.448, improving over the official baseline score of 0.429 (by 4.4%), with an overall mAP@50 of 0.505. A configuration study on the validation split further characterizes the impact of key training choices. Temporal robustness analysis reveals pronounced seasonality, with peak performance in February (mAP@50 = 0.610) under cooler conditions and degradation in April (mAP@50 = 0.438) during seasonal transition. Our findings suggest that drift-aware, grayscale-consistent augmentation is a straightforward and effective lever for improving long-term LWIR detection, while temporal diagnostics are essential for understanding failure modes in real deployments.

Despite progress in sign recognition, real-time systems that can robustly handle partial or uncertain visual cues remain limited. This challenge is especially pronounced when only a small set of keywords can be reliably detected, yet the system must still generate fluent, contextually appropriate natural-language output. In this paper, we present SignGPT, a real-time American Sign Language (ASL) keyword spotting and LLM-guided natural language generation system. We extract hand and upper-body landmarks and image crops using MediaPipe. We then employ a compact visual encoder with a BiLSTM/GRU temporal head trained with Connectionist Temporal Classification (CTC) to recognize fifteen ASL keywords without frame-level alignments. The detected keyword sequence (typically 2-5 keywords) conditions a large language model (LLM) to perform contextual sentence completion, returning the top three English sentence candidates consistent with the recognized keywords and a task prompt (e.g., domain constraints, style, or register). We outline an evaluation protocol that combines keyword-level sequence accuracy with sentence-level quality assessed via human preference studies. On a held-out test set, SignGPT achieves 97.0% sentence-level accuracy with LLM guidance versus 73.5% without, demonstrating that LLM-guided contextual completion effectively bridges real-time keyword spotting and coherent ASL-to-English sentence generation.

Multimodal biosensing is a promising route to reliable upper-limb intent detection for rehabilitation and humanrobot interaction. In the ABC 2026 Multimodal Upper-limb Movement Intent Detection Challenge (MUMIDC), synchronized EEG, EMG, and IMU recordings are provided together with trigger timestamps that delineate task stages. We propose a lightweight, trigger-aligned pipeline that (i) robustly parses raw MindRove CSV files (including corrupted/header-only trials), (ii) reconciles timestamp timebases and aligns multimodal streams to trigger events, (iii) extracts fixed-length statistical descriptors from cue and action windows, and (iv) compares unimodal and fusion strategies under two official protocols: Experiment 1 (within-subject) and Experiment 2 (leave-one-subject-out, LOSO). Across LOSO action decoding, EMG+IMU achieved 0.695 ± 0.208 accuracy and 0.694 ± 0.208 macro-F1, while lightweight late fusion with weight w=0.85 achieved 𝟎. 𝟔 𝟗 𝟔 ± 𝟎. 𝟐 𝟎 𝟖 accuracy and 0.694 ± 0.208 macro-F1. LOSO cue decoding was substantially more challenging; the best model (late fusion) reached 0.375 ± 0.101 accuracy and 0.362 ± 0.097 macro-F1. We analyze practical data issues encountered in the raw recordings and show that within-subject feature normalization and late fusion improve robustness, while cross-subject generalization remains the primary limitation for early-stage intention decoding.

With applications in animal monitoring, veterinary diagnostics, behavioral analysis, and robotics, real-time estimation of animal posture is an area of increasing interest in computer vision. In this work, we propose a method based on deep learning approaches to assess animal positions in real time. Selected for their applicability in both home and agricultural settings, the study centres on four animal classes: chicken, dog, horse, and cow. An important output of this work is a bespoke dataset created especially for posture estimation activities, including annotated videos. Every video in the dataset records continuous movement under different lighting and environmental contexts and runs for fifteen seconds. Keypoints marking important body joints for all types of animals were added to the extracted frames. Using the YOLOv8n posture architecture - which provides a balanced trade-off between speed and accuracy - we performed posture estimation. Although YOLO models are usually optimized for object recognition, we fine-tuned YOLOv8n-Pose to predict both bounding boxes and body keypoints, therefore enabling the real-time identification of intricate postural information. Trained on an annotated dataset using supervised learning, the model was tested on another test set from the same distribution. The proposed model achieves a PosePR mAP at the IoU threshold 0.5 of 99.5% in all classes, according to the experimental data. The dog class showed lower precision and F1 scores; the dog and horse classes showed decreased recall. The model maintains strong performance in real-time even if interclass posture variability and occlusion in video frames present natural difficulties. The system handles video input at an average frame rate enough for monitoring systems to be live. This study emphasizes the need for custom data sets tailored to real-world activities and the viability of employing YOLOv8n-Pose for the estimation of animal posture based on key points. Future directions include growing the data set, increasing keypoint accuracy, and including temporal consistency across frames. The data set is available at this link https://drive.google.com/drive/folders/1xci52bt9IxcYQrq36r2fQBaLvx3cSGHH#

Agriculture is one of the main economic pursuits of the 64 Bangladeshi districts.: in Bangladesh, about seventy percent of the workforce rely on agriculture for their living. Bangladesh's gross national revenue is much enhanced by the country's rice farming but attacks of insect pests have a great impact on rice harvests. Different insect pests require different management measures, so the accurate identification of paddy field insects is a crucial task that allows, e.g., the application of the appropriate poison for specific insect pests. This will also prevent the wasteful use of ineffective insecticides. The main challenge addressed in this work is to detect and instantly segment small harmful insects in paddy fields. To address this problem, we have used deep convolutional neural network (DCNN) learning, based on Mask-RCNN, therefore enabling a technique for visual localisation and classification of agricultural pest insects. We have also developed our own dataset of harmful insect annotated images. In the proposed Mask-RCNN model we used a ResNet101 backbone, which can detect and segment at the same time. The proposed model achieves an AP@0.5 of 85.7%, a mAP of 63.8%, and an AR@10 of 68.5, therefore generating an anticipated accuracy of 75%. ResNet101 performs better on all measures. The suggested approach should be able to identify and classify the small harmful insects, with suitable accuracy, in real-world deployments.

Transformer-based architectures have become the dominant approach for a wide array of machine learning tasks, including those in computer vision. Consequently, the prevalence of purely convolutional networks-particularly shallow-depth architectures for classification-has been in decline. In this work, we revisit Convolutional Neural Networks (CNNs) and propose a modern hybrid architecture that integrates Transformer-inspired components. Specifically, we introduce MLP Fusion, a model that incorporates Multi-Layer Perceptron (MLP) blocks, similar to those used in Vision Transformers, into CNN backbones prior to the classification stage. Additionally, we include intermediate 1 \times 1 convolutional layers within the backbone. This fusion is intended to enhance the representational capacity of CNNs by enriching their embedding space. Experimental evaluations on the CIFAR-10 and CIFAR-100 datasets show that MLP Fusion achieves better performance compared to compact CNN models reported in the literature.