MOBILE APPLICATION DEVELOPMENT FOR BLIND PEDESTRIANS TO PREVENT ROAD DANGERS

Abstract

According to the last statistic researches approximately more than one billion people worldwide live with some form of visual impairment. In turn, visual impairments limit people's ability to perform daily functions and affect their quality of life and ability to interact with the world around them. In the article mobile application development for blind pedestrians to prevent road dangers is presented. Short overviews of similar applications like Alexa, Via Opta Nav, and Object Detector are described. Each of described programs has disadvantages like limited use area, real-time object detection absence, use third-party or physical devices need. As a result, the main task of the study is to investigate modern hazard classification algorithms, improve the accuracy of the algorithm and develop software that will be able to identify hazards in real-time, does not require physical devices, and is operated using the simplest possible interface. For solving presented above problem solution based on MobiNetV2 and InceptionV3 open-source models for defining objects in a photo modification is presented. The presented solution consists of several steps like image input with further preprocessing, optimization and result processing. For the image input hosts receive data from the file system or local memory, perform any preprocessing, and then transmit the preprocessed data to the TPU cores. Preprocessing calls the parser, which in turn calls the parser function, where images are preprocessed. For the optimization stochastic gradient descent optimization and momentum optimizer are used. As a result, method of image classification for real-time hazard identification has been further developed. A model layer was developed that interprets the unbalanced results of the model and provides the necessary results to prevent accidents, which increased accuracy by 20%. A mobile application for road hazard recognition for blind pedestrians has been developed using the above. Presented results confirm the efficiency of the described approach. Also, described model and approach can be improved in further investigations.