Adaptive Object Detection for Indoor Navigation Assistance: A Performance Evaluation of Real-Time Algorithms
Abhinav Pratap, Sushant Kumar, Suchinton Chakravarty · 2025
Read Adaptive Object Detection for Indoor Navigation Assistance: A Performance Evaluation of Real-Time Algorithms by Abhinav Pratap, Sushant Kumar, Suchinton Chakravarty (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-18444.txt.
Introduction
Blind and low-vision people need indoor guides that spot obstacles fast without crying wolf. Abhinav Pratap, Sushant Kumar, and Suchinton Chakravarty compared four real-time detection algorithms for that job. The field covered two fast one-shot detectors and two heavier two-stage ones. An indoor objects dataset supplied the test scenes. Accuracy, speed, and fit to indoor settings were weighed together.
The results map the trade between precision and pace for each method. Heavier models read scenes more exactly but run slower on modest hardware. For a worn guide, late warnings are as bad as wrong ones. The authors offer their comparison as a selection guide for builders of navigation aids.
Catalog record
| CREATOR | Abhinav Pratap, Sushant Kumar, Suchinton Chakravarty |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-18444
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Adaptive Object Detection for Indoor Navigation Assistance: A Performance Evaluation of Real-Time Algorithms Abhinav Pratap, Sushant Kumar, Suchinton Chakravarty Abstract This study addresses the critical need for accurate and efficient object detection in assistive technologies for visually impaired individuals. We systematically evaluate the performance of four prominent real-time object detection algorithms—YOLO, SSD, Faster R-CNN, and Mask R-CNN—within the context of indoor navigation assistance. Our analysis, conducted on the Indoor Objects Detection dataset, focuses on key parameters including detection accuracy, processing speed, and adaptability to the unique challenges of indoor environments. The findings highlight the crucial trade-off between precision and processing efficiency, providing valuable insights for selecting optimal algorithms for real-time assistive navigation systems. This research contributes to a deeper understanding of adaptive machine learning applications that can significantly improve indoor navigation solutions for the visually impaired, promoting inclusivity and accessibility. I Introduction In today’s technology-driven society, there is an increasing emphasis on enhancing accessibility for visually impaired individuals. Indoor navigation poses unique challenges due to confined spaces, dynamic obstacles, and the need for precise and real-time object detection. While various object detection algorithms—such as YOLO [2, 8], SSD [2, 3], Faster R-CNN [7], and Mask R-CNN [7]—have been extensively studied in general computer vision contexts, their specific application to assistive indoor navigation remains underexplored. Recent studies have demonstrated the potential of object detection models in improving accessibility technologies. For example, Tinier-YOLO [11] and MobileNet-SSD [3] have shown promise in balancing accuracy and computational efficiency, making them viable candidates for resource-constrained environments. Similarly, research on assistive technologies for visually impaired individuals [13, 14] highlights the need for precise object detection in indoor settings to ensure safe and efficient navigation. However, these studies often lack a comprehensive evaluation of algorithm performance under the unique conditions of indoor environments, such as cluttered spaces and varying lighting conditions. This research aims to bridge this gap by systematically evaluating real-time object detection algorithms—YOLO, SSD, Faster R-CNN, and Mask R-CNN—on key parameters relevant to indoor navigation for visually impaired users. By focusing on the trade-offs between speed, accuracy, and adaptability, this study seeks to provide actionable insights for developing inclusive navigation systems that cater to the specific needs of this demographic. II Dataset The Indoor Objects Detection dataset was created with the aim of assisting persons with visual impairments in their day-to-day life. This dataset supports the ‘Object Detection for Blind People’ project under the AI Builders 2022 initiative. The project’s ultimate goal is the detection of indoor objects, emphasizing its relevance and applicability across various domains. The choice to utilize the Indoor Objects Detection dataset for testing our models in object detection stems from its intrinsic relevance to our research focus on enhancing navigation for individuals with visual impairments. This dataset specifically targets indoor environments, aligning seamlessly with the challenges faced by visually impaired persons in navigating confined spaces. By encompassing 7331 labeled objects across 10 different indoor classes, including doors, cabinets, and furniture, the dataset provides a distinct, and broad set of scenarios for testing the effectiveness of our models. The inclusion of bounding box annotations further facilitates precise object detection, crucial for developing navigation assistance systems [13, 14]. Leveraging this dataset allows us to tailor our models to the unique demands of indoor settings, fostering the creation of more accurate and adaptive solutions for real-time navigation assistance. The average areas for classes are shown in Fig. 1. The Co-Occurrence Matrix of class objects in the dataset makes this dataset more useful for our application. Fig. 1: Co-Occurrence Matrix of class object in Dataset III Related Work Object detection is a cornerstone of computer vision, and its applications have evolved significantly in recent years. YOLO, introduced as a real-time object detection model, revolutionized the field by achieving high detection speeds without compromising accuracy [2, 8]. Subsequent iterations, such as Tinier-YOLO, have optimized the model for constrained environments [11]. Similarly, SSD leverages predefined bounding boxes to balance speed and accuracy, making it a popular choice for dynamic settings [3]. Faster R-CNN and Mask R-CNN, while more computationally intensive, offer higher accuracy levels and detailed object segmentation, respectively. Studies like [6, 7] underscore their suitability for tasks requiring precision, but their applicability to real-time scenarios remains limited by their processing demands. Recent work on indoor navigation systems for visually impaired individuals [13, 14] highlights the critical role of object detection in addressing challenges such as identifying small objects and navigating crowded spaces. However, these studies often focus on a single algorithm or lack a comparative analysis across multiple models. Despite these advancements, existing research rarely explores the integration of these algorithms in assistive technologies tailored to indoor navigation. This gap is particularly pronounced in studies addressing real-time requirements and confined environments. Our work builds on these findings by conducting a comparative analysis of four prominent object detection models, emphasizing their performance in indoor scenarios and their implications for assistive navigation systems. The reviewed literature underscores significant progress in object detection and its applications. However, the unique challenges of indoor navigation—such as cluttered environments, dynamic obstacles, and the need for real-time responses—necessitate further investigation. By evaluating YOLO, SSD, Faster R-CNN, and Mask R-CNN under these specific conditions, this research addresses the limitations identified in prior studies and provides a robust framework for selecting algorithms that balance accuracy, speed, and adaptability. IV Algorithms IV-A YOLO (You Only Look Once) YOLO, standing for You Only Look Once, revolutionizes real-time object detection by employing a grid-based approach. The input image is split into an SxS grid, where every grid cell is tasked with directly forecasting bounding boxes and class probabilities. This innovative strategy eliminates the need for a multi-step process, enabling YOLO to achieve high speed and efficiency. The algorithm’s single-pass detection system contributes to its suitability for real-time applications, making it a favorite choice in various domains [2, 8]. Fig. 2: Single stage YOLO architecture A comprehensive working model of YOLO for a single frame is explained in Table I : TABLE I: Working of YOLO Steps Description Input Image Original image to be processed. Grid Division Image is divided into an S \times S grid. Each cell is responsible for object detection. Bounding Box Prediction Each grid cell predicts bounding boxes, class probabilities, and confidence scores. Non Maximum Suppression Redundant detections are filtered out using NMS based on confidence scores. Output Final bounding boxes, class probabilities, and confidence scores. While the explanation used a single input image for simplicity, YOLO is commonly used in a batch processing mode. Instead of processing images one by one, YOLO can efficiently process multiple frames simultaneously. This parallel processing allows for real-time detection in video streams by handling multiple frames concurrently [2, 8]. IV-B SSD (Single Shot Multibox Detector) SSD, or Single Shot MultiBox Detector, is an object detection technique optimized for a trade-off between accuracy and speed. It operates by employing a predefined set of boxes with various aspect ratios at each cell of the feature map. Simultaneously, it predicts both bounding box offsets and class scores. Let’s see the working of SSD in detail: • Feature Extraction: The input image is fed into a base Convolutional Neural Network (CNN), often based on architectures like VGG16. This CNN extracts feature maps