Wiper Arm Recognition and Counting Using Deep Learning Approach

Ling, Hua Jian (2021) Wiper Arm Recognition and Counting Using Deep Learning Approach. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Ling Hua Jian.pdf Restricted to Registered users only Download (9MB)

Abstract

Malaysia's Government promotes the transformation of industry 4.0 in recent years. The program leads most of the industry to begin to transform their manufacturing towards automation. A wiper arm recognition and counting system will be designed to fulfil industry demands since it is not available in the market yet. This system's primary goal is to reduce workforce and human error during the verifying process by replacing the human's naked eye with the wiper arm recognition and counting system. The accuracy of previous work for wiper arm detection is 71.92%. In this thesis, the proposed deep learning approach for the wiper arms recognition is a one-stage object detection algorithm named YOLOv4. While TensorFlow API was used to count the number of wipers in each model. Therefore, YOLOv4 was implemented in TensorFlow to perform object detection and counting simultaneously. The YOLOv4 model was trained well and able to predict up to 4 types of wiper model. In a nutsell, the mean average precision (mAP) and F1-Score achieved by this system are 100% without any false detection and misprediction. The system implemented on YOLOv4 also able to predict well on several conditions such as object occlusion, different scales of object, different light conditions.