Dynamic Object Prioritization and Contextual Scene Analysis in Assistive Navigation using a Hybrid Self-Attentive Convolutional Bi-GRU-Net Ar-chitecture

Abstract

The provision of autonomous mobility for the visually impaired necessitates the integration of high-fidelity en-vironmental perception and intelligent hazard assessment. This paper presents a novel deep learning paradigm, ‘Dynamic object prioritization and contextual scene analysis in assistive navigation using a hybrid self attentive convolutional Bi-GRU-Net architecture’, engineered to enhance scene interpretation through hierarchical feature extraction. The research addresses the critical need for systems that can not only identify obstacles but also pri-oritize them based on navigational relevance. The proposed framework utilizes a systematic pipeline beginning with the acquisition of complex environmental data from the SUN RGB-D dataset. To ensure signal integrity, raw inputs undergo preprocessing via Modified Mean Filtering (Mod_MFil) for noise attenuation, followed by a hybrid segmentation strategy. This strategy leverages the Watershed (WS) method and Mean-Shift (MS) clustering to precisely delineate semantic regions of interest. At its core, the SA-Conv-DSBiGR Net facilitates a sophisticated fusion of spatial and temporal features, utilizing self-attention mechanisms to weigh critical visual cues. Furthermore, an innovative prioritization module is introduced, employing the Minimum Redundancy Maximum Relevance (MRMR) model. By calculating inter-object relationships through Cosine similarity and Euclidean distance metrics, the system dynamically ranks objects to provide actionable feedback. Empirical evaluations conducted in a Python environment demonstrate that the architecture achieves a peak accuracies of 80.81%, 77.30%, and 78.20%, and precision scores of 86.30%, 94.22%, and 82.16% for chairs, dogs, and pot plants, respectively. The system's diagnostic reliability is further evidenced by a ROC AUC value of 99.64%, indicating nearly perfect predictive performance. significantly outperforming baseline models and establishing a new benchmark for assistive computer vision technologies..