MODIFICATION OF EFFICIENTDET-LITE MODEL ALGORITHM USING MINI-BIFPN AND NMS-DIOU FOR ROAD OBJECT DETECTION ON ANDROID SMARTPHONE DEVICES

• Main Author: Rahimin Ramadhani, Khairur
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85506
• Institution: Institut Teknologi Bandung

Smartphone technology has rapidly developed the potential to be used as an alternative to surveillance cameras or dashcams in monitoring and detecting vehicle objects on the highway. However, with the limited hardware and varied specifications of smart phones, a vehicle detection deep learning model is needed that is not only accurate but also lightweight to run. One of the popular object detection model algorithms is EfficientDet developed by Google (Tan et al., 2020). The EfficientDet model also has an EfficientDet-Lite variant that is intended for use on mobile devices. To improve the performance of the model, the authors conducted experiments on their own dataset of highway conditions to modify the BiFPN and Detection Head (NMS) networks in EfficientDet-Lite, and obtained a Mini EfficientDet-Lite NMS-DIoU model that produces mAP of 9.20%. There is a 3.30% increase in mAP from the EfficientDet-lite baseline model.