International Journal of Image Processing and Smart Sensors

Multi Modal Fusion Framework of Thermal Imaging and Low Resolution LiDAR for Pedestrian Detection in Adverse Weather

2026-10-03T10:49:54+00:00

Background: Adverse weather conditions (fog, rain, snow, low light) severely degrade RGB and high‑resolution LiDAR-based pedestrian detection, compromising safety in autonomous driving and surveillance. Low‑resolution (4–16 beam) LiDAR is cost‑effective and privacy‑preserving but suffers from extreme spatial sparsity, while thermal imaging provides weather‑agnostic dense texture yet lacks geometric precision.

Purpose: This paper proposes a novel multi‑modal fusion framework that synergistically integrates thermal infrared imagery with low‑resolution smart LiDAR to achieve robust pedestrian detection under adverse weather.

Methods: The framework comprises a lightweight thermal CNN for heat distribution and edge extraction, a PointPillars‑style sparse convolution branch for LiDAR point clouds, and a cross‑attention fusion module augmented with a learned mask‑guided interpolation to densify sparse LiDAR features. Experiments are conducted on augmented KAIST and FLIR datasets with physics‑based fog, rain, and snow, plus real‑world validation.

Findings: The proposed method achieves 89.3% mean average precision (mAP@0.5) across all adverse conditions, outperforming thermal‑only by 15.3% and LiDAR‑only by 24.6%. Under dense fog (visibility <50 m), it retains 88.3% mAP, a mere 4.1% drop from clear weather. The system runs at 50 fps on an RTX 3090 and 25 fps on a Jetson Orin after INT8 quantization.

Conclusion: Fusing low‑resolution LiDAR with thermal imaging effectively compensates for each modality’s weaknesses, enabling all‑weather pedestrian detection at low sensor cost. Recommendation: Deploy the framework in cost‑sensitive ADAS, intelligent surveillance, and outdoor robotics where all‑weather reliability and privacy are essential. Future integration with radar or federated learning is suggested.

A Review on the Emergence of Intelligent Oncology by Sensor-Guided Drug Delivery

2026-09-19T05:55:57+00:00

For decades, the fight against cancer has been a brutal war of attrition. Chemotherapy, medicine’s necessary evil, operates like a scorched-earth campaign—highly effective, yet devastatingly indiscriminate, destroying both the invasive enemy cells and the healthy host tissue. Conventional cancer chemotherapy often suffers from limitations such as non-specific biodistribution, systemic toxicity, and inadequate drug concentrations at tumor sites, leading to suboptimal therapeutic outcomes and severe side effects. This study presents the transformative potential of integrating advanced sensor technology with drug delivery systems to address these critical oncology challenges. By enabling real-time monitoring and responsive drug release, sensor-enabled platforms promise to revolutionize cancer treatment through enhanced precision, personalization, and efficacy. This study explores innovative approaches where miniaturized biosensors (e.g., electrochemical, optical, pH, temperature, or biomarker-responsive) are incorporated into smart drug carriers (e.g., nanoparticles, hydrogels, implantable microdevices). These intelligent systems are designed to sense specific tumor microenvironmental cues (e.g., abnormal pH, hypoxia, enzyme overexpression) or patient physiological parameters, subsequently triggering the controlled and localized release of anticancer agents. This feedback-controlled mechanism ensures targeted delivery, minimizes off-target accumulation, and allows for on-demand dosing adjustments, thereby maximizing therapeutic effects while significantly reducing systemic toxicity. The integration of sensor technology ushers in an era of truly personalized cancer therapy, offering hope for improved patient quality of life and more effective disease management.

Automated Handwritten Text Analysis and Recognition Using Neural Networks

2026-09-25T05:10:34+00:00

Handwritten text recognition (HTR) is a difficult problem in Optical Character Recognition (OCR) systems, owing to the extreme variability of handwriting styles, degradation of image quality, and diverse character appearances. The present paper proposes a comprehensive framework based on deep learning techniques, which incorporates Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) models, and Connectionist Temporal Classification (CTC) for effective handwritten text recognition. The hybrid model utilizes CNNs to learn spatial features from handwritten text images, BiLSTMs to learn sequential relationships between characters, and CTC to learn sequences directly from images without requiring explicit character segmentation. The model incorporates various preprocessing techniques to improve its performance, including image normalization, noise reduction, and data augmentation, to achieve robustness against diverse handwriting styles. The performance of the model has been evaluated on benchmark datasets, achieving state-of-the-art performance with a character error rate of 2.95% and a word error rate of less than 7%. The model achieves an accuracy of 94% during validation, with efficient computation to suit practical scenarios. The present study contributes to the advancement of automated document digitization, accessibility for the visually impaired, efficient data processing in banking, healthcare, and education, among other areas.