OPTIMIZED DEEP CNN BASED OBSTACLE DETECTION FOR AIDING VISUALLY IMPAIRED PERSONS

A visually impaired person faces several challenges while they are moving towards unfamiliar environments. Hence, object detection approaches provide a major solution for this issue. For that, various researchers have developed obstacle detection approaches to help blind people however they have certain limitations. In this research, the optimized deep learning techniques, named Social Optimization algorithm (SOA)-based Deep Convolutional Network (Deep CNN) is developed for assisting visually damaged persons. For effective obstacle detection, the input videos are converted to multiple frames. In feature extraction, relevant features, such as Convolutional Neural Network (CNN) features, Shape Local Binary Texture (SLBT), and hierarchical skeleton features are extracted for further processing. Moreover, the object detection process is carried out using Generative Adversarial Network (GAN). In addition, the object recognition process is done by Deep CNN in which all the layers of Deep CNN are trained using SOA. In addition, the experimental result demonstrates that the developed model attained the testing accuracy, mean average precision (mAP), and recall values of 0.9485, 0.9596, and 0.9735, correspondingly. KEYWORDS: Deep CNN, Social Optimization algorithm (SOA), Generative Adversarial Network (GAN), Obstacles detection Una persona con discapacidad visual se enfrenta a varios desafios mientras se desplaza hacia entornos desconocidos. Por lo tanto, los enfoques de deteccion de objetos proporcionan una solution importante para este problema. Para eso, varios investigadores han desarrollado enfoques de deteccion de obstaculos para ayudar a las personas ciegas, sin embargo, tienen ciertas limitaciones. En esta investigation, las tecnicas optimizadas de aprendizaje profundo, denominadas Red Convolucional Profunda basada en el algoritmo de Optimizacion Social (SOA) (Deep CNN) se desarrollan para ayudar a las personas con discapacidad visual. Para una deteccion efectiva de obstaculos, los videos de entrada se convierten en multiples cuadros. En la extraction de caracteristicas, las caracteristicas relevantes, como las caracteristicas de la red neuronal convolucional (CNN), la textura binaria local de forma (SLBT) y las caracteristicas del esqueleto jerarquico se extraen para su posterior procesamiento. Ademas, el proceso de deteccion de objetos se lleva a cabo utilizando Generative Adversarial Network (GAN). Ademas, el proceso de reconocimiento de objetos lo reali