An End-to-End Object Detection System in Indoor Environments Using Lightweight Neural Network

Vision plays a pivotal role in how humans interact with and understand their surroundings.Despite significant advancements in assistive technology, the specific challenges of indoor navigation for the visually impaired and blind (VIP) community remain inadequately addressed.Recognizing and locating objects within indoor environments is crucial for enhancing the independence and safety of VIP individuals.This paper introduces an innovative deep-learning framework designed for superior object detection in indoor settings, alongside a comprehensive new dataset tailored for this purpose.At the heart of the proposed system is an optimized version of the YOLOv3 architecture, re-engineered to significantly reduce computational demands while maintaining high accuracy.The newly developed indoor dataset encompasses over 11,000 images featuring 25 essential object categories, curated to represent a variety of lighting conditions and complex indoor scenarios that pose navigational hazards.This dataset not only challenges the model with real-world complexities but also facilitates the training of more robust, efficient neural networks.Experimental results underscore the proposed model's exceptional performance, demonstrating marked improvements in detection precision and system efficiency, thereby offering a promising new direction for assistive technologies in indoor navigation.The obtained results of the proposed indoor objects detection system are highly promising, showcasing a mean average precision (mAP) of 89.78% across all 25 object classes.This impressive performance indicates the system's robust capability in accurately detecting a diverse range of objects commonly found in indoor environments.