Balasaheb Agarkar
Research Subjects
Balasaheb Agarkar has a total of 9 co-authors. Here are the top co-authors by number of shared publications:
- Sachin Vasant Chaudhari
Savitribai Phule Pune University
4 shared publications - Rekha Labade
Department of Electronics and Telecommunication Engineering, Sanjivani College of Engineering, Kopargaon, India
2 shared publications - Sachin Mandlik
Savitribai Phule Pune University
2 shared publications Rekha Labade
1 shared publicationSachin Chaudhari
1 shared publicationSachin Mandlik
1 shared publication
Latest Publications
- Open Access
- 8 May 2026
- Published by Universitat Autonoma de Barcelona in ELCVIA Electronic Letters on Computer Vision and Image Analysis
- Vol. 25 (2) , 79-105
Dual Attention Embedded Inception Framework for Robust Biometric Gait Recognition
Gait recognition has emerged as an important biometric modality due to its non-invasive nature and suitability for surveillance and security applications. However, achieving robustness under real-world variations in viewpoint, clothing, and carrying conditions remains a significant challenge. This paper introduces \textbf{AttIncGait}, a deep learning framework that integrates Inception-based multi-scale feature extraction with dual-path attention for effective gait recognition. Unlike prior methods that treat attention as an auxiliary or late-stage refinement, our approach embeds spatial and channel attention directly within Inception modules, enabling simultaneous multi-scale representation and adaptive relevance weighting. This structural integration enhances discriminative capability while preserving computational efficiency. Experiments on CASIA-B and OU-MVLP datasets demonstrate state-of-the-art performance: 97.5% accuracy on OU-MVLP and a 2.6% improvement over the best existing method under clothing variation in CASIA-B. Ablation studies further reveal that spatial and channel attention individually improve accuracy, while their joint integration yields an overall +8.5% gain on OU-MVLP. These results validate the effectiveness of attention-driven multi-scale fusion for gait recognition and highlight the potential of AttIncGait for real-world biometric identification and mobility analysis.
- 8 May 2026
- Published by Universitat Autonoma de Barcelona in ELCVIA Electronic Letters on Computer Vision and Image Analysis
- Vol. 25 (2) , 79-105
Open AccessDual Attention Embedded Inception Framework for Robust Biometric Gait Recognition
Gait recognition has emerged as an important biometric modality due to its non-invasive nature and suitability for surveillance and security applications. However, achieving robustness under real-world variations in viewpoint, clothing, and carrying conditions remains a significant challenge. This paper introduces \textbf{AttIncGait}, a deep learning framework that integrates Inception-based multi-scale feature extraction with dual-path attention for effective gait recognition. Unlike prior methods that treat attention as an auxiliary or late-stage refinement, our approach embeds spatial and channel attention directly within Inception modules, enabling simultaneous multi-scale representation and adaptive relevance weighting. This structural integration enhances discriminative capability while preserving computational efficiency. Experiments on CASIA-B and OU-MVLP datasets demonstrate state-of-the-art performance: 97.5% accuracy on OU-MVLP and a 2.6% improvement over the best existing method under clothing variation in CASIA-B. Ablation studies further reveal that spatial and channel attention individually improve accuracy, while their joint integration yields an overall +8.5% gain on OU-MVLP. These results validate the effectiveness of attention-driven multi-scale fusion for gait recognition and highlight the potential of AttIncGait for real-world biometric identification and mobility analysis.
- Open Access
- 7 December 2025
- Published by Springer Nature in Journal of Engineering and Applied Science
- Vol. 72 (1) , 234
Pose-invariant gait recognition using an enhanced inception-ResNet and vision transformer framework
Gait recognition is an emerging biometric technique with wide-ranging applications in surveillance, security, and healthcare. Despite its advantages over traditional biometric modalities, performance remains sensitive to factors such as viewpoint variations, occlusions, and posture changes. To address these challenges, this study proposes a hybrid gait recognition framework that integrates Inception-ResNet and Vision Transformer (ViT) architectures for robust, pose-invariant feature extraction. The framework employs OpenPose to estimate skeletal keypoints, which are processed through a dual-path network: the Inception-ResNet branch captures multi-scale spatial representations of local gait patterns, while the ViT branch models long-range temporal dependencies via self-attention. This combination enables a comprehensive representation of both fine-grained spatial features and global temporal dynamics. Experiments conducted on benchmark datasets—CASIA-B, OU-MVLP, GREW and Gait3D—demonstrate that the proposed method surpasses state-of-the-art approaches, including GaitSet, GaitPart, and GaitGL, particularly under cross-view and occluded conditions. The framework achieves Rank-1 accuracy of 98.2% on CASIA-B under normal walking, 92.7% on OU-MVLP, 85.4% on GREW and 79.4% on Gait3D highlighting its robustness and practical potential for real-world applications.
- 7 December 2025
- Published by Springer Nature in Journal of Engineering and Applied Science
- Vol. 72 (1) , 234
Open AccessPose-invariant gait recognition using an enhanced inception-ResNet and vision transformer framework
Gait recognition is an emerging biometric technique with wide-ranging applications in surveillance, security, and healthcare. Despite its advantages over traditional biometric modalities, performance remains sensitive to factors such as viewpoint variations, occlusions, and posture changes. To address these challenges, this study proposes a hybrid gait recognition framework that integrates Inception-ResNet and Vision Transformer (ViT) architectures for robust, pose-invariant feature extraction. The framework employs OpenPose to estimate skeletal keypoints, which are processed through a dual-path network: the Inception-ResNet branch captures multi-scale spatial representations of local gait patterns, while the ViT branch models long-range temporal dependencies via self-attention. This combination enables a comprehensive representation of both fine-grained spatial features and global temporal dynamics. Experiments conducted on benchmark datasets—CASIA-B, OU-MVLP, GREW and Gait3D—demonstrate that the proposed method surpasses state-of-the-art approaches, including GaitSet, GaitPart, and GaitGL, particularly under cross-view and occluded conditions. The framework achieves Rank-1 accuracy of 98.2% on CASIA-B under normal walking, 92.7% on OU-MVLP, 85.4% on GREW and 79.4% on Gait3D highlighting its robustness and practical potential for real-world applications.
- Open Access
- 1 August 2025
- Published by Poltekkes Kemenkes Surabaya in Journal of Electronics, Electromedical Engineering, and Medical Informatics
- Vol. 7 (4) , 1060-1073
Unified Deep Architectures for Real-Time Object Detection and Semantic Reasoning in Autonomous Vehicles
The development of autonomous vehicles (AVs) has revolutionized the transportation industry, promising to boost mobility, lessen traffic, and increase safety on roads. However, the complexity of the driving environment and the requirement for real-time processing of vast amounts of sensor data present serious difficulties for AV systems. Various computer vision approaches, such as object detection, lane detection, and traffic sign recognition, have been investigated by researchers in order to overcome these issues. This research presents an integrated approach to autonomous vehicle perception, combining real-time object detection, semantic segmentation, and classification within a unified deep learning architecture. Our approach leverages the strengths of existing frameworks, including MultiNet’s real-time semantic reasoning capabilities, the fast-encoding methods of PointPillars to identify objects from point clouds, as well as the reliable one-stage monocular 3D object detection system. The offered model tries to improve computational efficiency and accuracy by utilizing a shared encoder and task-specific decoders that perform classification, detection, and segmentation concurrently. The architecture is evaluated against challenging datasets, illustrating outstanding achievements in terms of speed and accuracy, suitable for real-time applications in autonomous driving. This integration promises significant advancements in the perception systems of autonomous vehicles a providing in-depth knowledge of the vehicle’s environment through efficient concepts of deep learning techniques. In our model, we used Yolov8, MultiNet, and during training got accuracy 93.5%, precision 92.7 %, recall 82.1% and mAP 72.9%.
- 1 August 2025
- Published by Poltekkes Kemenkes Surabaya in Journal of Electronics, Electromedical Engineering, and Medical Informatics
- Vol. 7 (4) , 1060-1073
Open AccessUnified Deep Architectures for Real-Time Object Detection and Semantic Reasoning in Autonomous Vehicles
The development of autonomous vehicles (AVs) has revolutionized the transportation industry, promising to boost mobility, lessen traffic, and increase safety on roads. However, the complexity of the driving environment and the requirement for real-time processing of vast amounts of sensor data present serious difficulties for AV systems. Various computer vision approaches, such as object detection, lane detection, and traffic sign recognition, have been investigated by researchers in order to overcome these issues. This research presents an integrated approach to autonomous vehicle perception, combining real-time object detection, semantic segmentation, and classification within a unified deep learning architecture. Our approach leverages the strengths of existing frameworks, including MultiNet’s real-time semantic reasoning capabilities, the fast-encoding methods of PointPillars to identify objects from point clouds, as well as the reliable one-stage monocular 3D object detection system. The offered model tries to improve computational efficiency and accuracy by utilizing a shared encoder and task-specific decoders that perform classification, detection, and segmentation concurrently. The architecture is evaluated against challenging datasets, illustrating outstanding achievements in terms of speed and accuracy, suitable for real-time applications in autonomous driving. This integration promises significant advancements in the perception systems of autonomous vehicles a providing in-depth knowledge of the vehicle’s environment through efficient concepts of deep learning techniques. In our model, we used Yolov8, MultiNet, and during training got accuracy 93.5%, precision 92.7 %, recall 82.1% and mAP 72.9%.
- Open Access
- 11 February 2025
- Published by Poltekkes Kemenkes Surabaya in Journal of Electronics, Electromedical Engineering, and Medical Informatics
- Vol. 7 (2) , 450-459
Deep Learning Approach for Segmenting Nuchal Translucency Region in Fetal Ultrasound Images for Detecting Down Syndrome using GoogLeNet and AlexNet
Down syndrome (DS) is a chromosomal disorder linked to intellectual impairment and developmental delays in babies. The primary prenatal indicator for detecting DS during the initial stages of gestation is the thickness of nuchal translucency (NT). This paper introduces a GoogLeNet model based on convolutional neural networks (CNN) for the semantic segmentation of the NT region from ultrasound fetal images, facilitating rapid and cost-effective diagnosis in the early stages of the gestational period. A transfer learning methodology with AlexNet is employed to train the NT regions for the detection of DS. The Inception module of GoogLeNet enables the model to simultaneously capture characteristics at various sizes of images. The capacity to extract both intricate and broad characteristics can improve the model’s performance in precisely identifying the NT area. This will function as an exceptional tool for physicians in screening of DS, enhancing the detection rate and providing a substantial opinion for early diagnosis. The proposed deep learning approach attained an accuracy of 96.18% and Jaccard index of 0.967 for NT region segmentation utilizing GoogLeNet. A confusion matrix was used to evaluate the image classification by AlexNet model's effectiveness, and the results showed an overall accuracy of 97.84%, ROC-AUC of 98.45%, recall of 99.64%, precision of 96.04%, and F1 score of 97.80%. The proposed deep learning method produced remarkable outcomes and can be applied to the identification of DS in medical field. This method identifies individuals at increased risk for this condition and enables termination in the early stages of pregnancy.
- 11 February 2025
- Published by Poltekkes Kemenkes Surabaya in Journal of Electronics, Electromedical Engineering, and Medical Informatics
- Vol. 7 (2) , 450-459
Open AccessDeep Learning Approach for Segmenting Nuchal Translucency Region in Fetal Ultrasound Images for Detecting Down Syndrome using GoogLeNet and AlexNet
Down syndrome (DS) is a chromosomal disorder linked to intellectual impairment and developmental delays in babies. The primary prenatal indicator for detecting DS during the initial stages of gestation is the thickness of nuchal translucency (NT). This paper introduces a GoogLeNet model based on convolutional neural networks (CNN) for the semantic segmentation of the NT region from ultrasound fetal images, facilitating rapid and cost-effective diagnosis in the early stages of the gestational period. A transfer learning methodology with AlexNet is employed to train the NT regions for the detection of DS. The Inception module of GoogLeNet enables the model to simultaneously capture characteristics at various sizes of images. The capacity to extract both intricate and broad characteristics can improve the model’s performance in precisely identifying the NT area. This will function as an exceptional tool for physicians in screening of DS, enhancing the detection rate and providing a substantial opinion for early diagnosis. The proposed deep learning approach attained an accuracy of 96.18% and Jaccard index of 0.967 for NT region segmentation utilizing GoogLeNet. A confusion matrix was used to evaluate the image classification by AlexNet model's effectiveness, and the results showed an overall accuracy of 97.84%, ROC-AUC of 98.45%, recall of 99.64%, precision of 96.04%, and F1 score of 97.80%. The proposed deep learning method produced remarkable outcomes and can be applied to the identification of DS in medical field. This method identifies individuals at increased risk for this condition and enables termination in the early stages of pregnancy.
- Open Access
- 1 February 2025
- Published by Institute of Advanced Engineering and Science in International Journal of Electrical and Computer Engineering (IJECE)
- Vol. 15 (1) , 349-355
Review of gait recognition systems: approaches and challenges
Gait recognition (GR) has emerged as a significant biometric identification technique, leveraging an individual's walking pattern for various applications such as surveillance, forensic analysis, and person identification. Despite its non-intrusive nature, GR systems face challenges due to their sensitivity to pose variations, limiting functionality in real-world scenarios where people exhibit diverse walking styles and body orientations. This review paper aims to comprehensively discuss GR systems, focusing on approaches and challenges in designing accurate and robust systems capable of handling bodily variations. GR's prominence spans across domains including surveillance, security, healthcare, and human-computer interaction, positioning it as a versatile biometric modality complementary to the traditional methods like fingerprint and face recognition. The review offers an in-depth analysis of GR systems, detailing silhouette-based, model-based, and deep-learning approaches. Silhouette-based methods capture gait information by analyzing the outline and locomotion of a person’s silhouette, while model-based approaches utilize skeletal models to describe gait patterns. The paper elucidates the challenges and limitations of GR systems, encompassing factors such as walking conditions, clothing, viewpoint, and environmental influences. Additionally, it explores potential future directions in GR research, highlighting the technology’s ongoing evolution and integration into diverse applications. As a valuable resource, this review serves researchers, practitioners, and policymakers by providing insights into the current state of GR systems and avenues for further research and development. It underscores the importance of addressing challenges to enhance GR’s accuracy and robustness, ensuring its continued relevance in biometric identification across various domains.
- 1 February 2025
- Published by Institute of Advanced Engineering and Science in International Journal of Electrical and Computer Engineering (IJECE)
- Vol. 15 (1) , 349-355
Open AccessReview of gait recognition systems: approaches and challenges
Gait recognition (GR) has emerged as a significant biometric identification technique, leveraging an individual's walking pattern for various applications such as surveillance, forensic analysis, and person identification. Despite its non-intrusive nature, GR systems face challenges due to their sensitivity to pose variations, limiting functionality in real-world scenarios where people exhibit diverse walking styles and body orientations. This review paper aims to comprehensively discuss GR systems, focusing on approaches and challenges in designing accurate and robust systems capable of handling bodily variations. GR's prominence spans across domains including surveillance, security, healthcare, and human-computer interaction, positioning it as a versatile biometric modality complementary to the traditional methods like fingerprint and face recognition. The review offers an in-depth analysis of GR systems, detailing silhouette-based, model-based, and deep-learning approaches. Silhouette-based methods capture gait information by analyzing the outline and locomotion of a person’s silhouette, while model-based approaches utilize skeletal models to describe gait patterns. The paper elucidates the challenges and limitations of GR systems, encompassing factors such as walking conditions, clothing, viewpoint, and environmental influences. Additionally, it explores potential future directions in GR research, highlighting the technology’s ongoing evolution and integration into diverse applications. As a valuable resource, this review serves researchers, practitioners, and policymakers by providing insights into the current state of GR systems and avenues for further research and development. It underscores the importance of addressing challenges to enhance GR’s accuracy and robustness, ensuring its continued relevance in biometric identification across various domains.