Moving Object Detection Based on Convolutional Neural Network

Abstract

In recent years self-driving technologies have become a popular research trend No matter for the current advanced assisted driving system (ADAS) or the future self-driving car the smart detections are needed to assist drivers to avoid any accidents on the road In particular the detections of moving objects in various conditions of road driving are relatively difficult and important In this thesis we proposed a moving object detection system for ADAS based on convolutional neural networks The reduction of the memory usage and computation load will be another concern such that the system can be appled in lightweight devices We also proposed a focal loss as the objective function to improve the training efficiency The designed detector mainly recognizes three targets: pedestrians riders and cars that are common on the roads In order to be more suitable for the circumstances in Taiwan we also collected and constructed dataset on Taiwanese urban roads especially the moto-riders data because the number of motorcycles in Taiwan is much larger than those in other developed countries The experimental results demonstrate that our proposed system achieves reasonable accuracy and keeps a good balance between detection rates and false alarm rates

Date of Award	2018 Aug 3
Original language	English
Supervisor	Jar-Ferr Yang (Supervisor)