Real-time object recognition and accurate environmental scene detection are critical challenges for developing effective navigational assistance systems for the visually impaired. Given that millions globally suffer from severe visual impairments, efficient and precise computer vision solutions are paramount. This paper introduces a novel deep learning framework, the Self-Attentive Convolutional Dual Stacked Bidirectional Gated Recurrent Network (SA-Conv-SBiGR Net), specifically engineered for robust object detection within complex scenes. The proposed methodology employs a sequential processing pipeline that begins with data acquisition from the SUN RGB-D dataset. Initial data refinement is performed using image resizing in conjunction with the Modified Mean Filtering (Mod_MFil) technique for enhanced noise suppression. The process continues with a two-stage segmentation approach utilizing both the Watershed Method (WS) and the Mean-Shift (MS) model to accurately delineate regions of interest. Feature classification is then handled by the SA-Conv-DSBiGRNet, which leverages its self-attention and dual-stacked architecture for superior spatial and temporal feature learning. Crucially, the framework integrates a priority-ranking mechanism based on the Minimum Redundancy Maximum Relevance (MRMR) model. Object prioritization is dynamically determined by evaluating inter-object relationships using Cosine similarity and Euclidean distance scores. Simulation results, conducted using Python and the SUN RGB-D dataset, validate the framework's efficacy. The proposed model demonstrates superior performance over existing methodologies, achieving a peak accuracy of 82.3%, confirming its potential as a highly reliable tool for visual assistance technology.