ABUAD Journal of Engineering Research and Development (AJERD) ISSN (online): 2645-2685; ISSN (print): 2756-6811 Volume 5, Issue 1, 147-154 www.ajerd.abuad.edu.ng/ 147 Development of a Convolutional Neural Network-Based Object Recognition System for Uncovered Gutters and Bollards Ibrahim Adepoju ADEYANJU *1 , Muhammed Adekunle AZEEZ 2 , Oluwaseyi Olawale BELLO 3 , Taofeeq Alabi BADMUS 4 and Mayowa Oyedepo OYEDIRAN 5 1,2,4 Department of Computer Engineering, Federal University, Oye-Ekiti, Nigeria ibrahim.adeyanju@fuoye.edu.ng;azmudray1@gmail.com;taofeeq.badmus@fuoye.edu.ng 3 Department of Computer Engineering, Ekiti State University, Ado-Ekiti, Ekiti State, Nigeria bello.oluwaseyi@eksu.edu.ng 5 Department of Computer Engineering, Ajayi Crowther University, Oyo, Oyo State, Nigeria mo.oyediran@acu.edu.ng Corresponding Author: ibrahim.adeyanju@fuoye.edu.ng, +2348132876689 Date Submitted: 09/02/2022 Date Accepted: 21/06/2022 Date Published: 30/06/2022 Abstract: Machine learning and deep learning have advanced considerably over the last few years with machine intelligence transitioning from laboratory to several industrial applications. Among the deep learning techniques, Convolutional Neural Networks (CNN) have been shown to have one of the best performances in image recognition. CNN has been used for the recognition of a lot of outdoor objects such as buildings, potholes, and cars but with little attention to the recognition of uncovered gutters and bollards, typically found in urban areas and higher institution environments of most developing countries. Hence, a CNN-based object recognition system for uncovered gutters and bollards, with high accuracy and low time complexity, was developed in this research. This can be used to aid outdoor navigation for the visually impaired. The images of uncovered gutters and bollards were captured locally with a high-resolution camera. The datasets were pre-processed by resizing the images and annotations carried out to generat e the images’ textual equivalent as well as define specific object boundaries. CNN was applied for feature extraction and recognition with two convolutional layers, two pooling layers, and a fully connected layer. The system implementation was done with Python programming language, OpenCV libraries, and Yolov4 as the CNN version with a percentage split experimental evaluation methodology. Results from experiments on the uncovered gutter dataset gave accuracy and average computational testing time of 80% and 0.4 s, respectively. Similarly, the bollards dataset with multiple bollards per image gave accuracy and average computational testing time of 72% and 0.47 s, respectively. The output of this research will be useful for outdoor navigation of the visually impaired when integrated into appropriate electronic hardware. Keywords: Deep learning, convolutional neural network, object recognition, gutter, bollard. 1. INTRODUCTION The emergence of new technology has tremendously contributed to the way we walk, talk, see and interact in the environment. One of the examples of these technologies is the assistive technology built to support people living with disabilities. Mobility aids such as wheelchairs, scooters, walkers, canes, crutches [1] are available to assist people with physical disabilities [2]. Globally, the number of people of all ages with visual impairments is estimated to be 285 million, of whom 39 million are blind [3]. About 26.3 million people in Africa have a form of visual impairment, 20.4 million have low vision, and 5.9 million are estimated to be blind [4]. Visual impairment can limit people’s ability to perform everyday tasks and affect the quality of life and ability to interact with the surrounding world [5]. There are several traditional methods of assistance provided for the visually impaired. These methods include various types of canes that help them in navigation. However, the cane is insufficient because every day routine mobility requires mastery of a regular route, and this requires memorization of the paths for a walk around. Guide dogs are available to lead visually impaired people around obstacles. The dogs have been trained to guide, but they are red- green colour blind and incapable of interpreting street signs [6]. Assistive technologies have been developed to address many of these challenges. Computer software and hardware, such as voice recognition programs, screen readers, and screen enlargement applications, to assist people with movement and sensory impairments use computers and mobile devices [7]. Technological infrastructure aids visually impaired navigation, such as voice announcements on buses and ‘talking’ crosswalks for outdoor navigation and Braille signs inside buildings. However, the devices are not universally available.