Deep learning for mobile phone detection while driving by image recognition
Majority of the existing works are based on in-car camera system mounted inside the vehicles, which is more suitable to be implemented as driver assistance system. This is because most of the driver assistance systems only are available for advanced cars. Besides that, a large amount of training dat...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2020
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/93010/1/YeohMeiHweiMSKE2020.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Majority of the existing works are based on in-car camera system mounted inside the vehicles, which is more suitable to be implemented as driver assistance system. This is because most of the driver assistance systems only are available for advanced cars. Besides that, a large amount of training datasets is required to train the driver monitoring system. However, there are no existing datasets that are captured based on camera system mounted outside the vehicles, thus the total amount of dataset acquired is limited for this project. Therefore, the aim of conducting this thesis is to develop a camera-based automated image recognition system that is mounted outside the vehicles for detecting the driver using mobile phones for calling while driving. Since there are no datasets available for this project, the images are captured and collected using Fujifilm XT 10 with XF 35mm f2 lens at overhead bridge nearby Spice Arena, Penang, Malaysia. The captured time is from 5pm to 7.30 pm. There are a total of 2,340 images are captured and collected. However, about 42 % of the captured images are discarded, left with1,348 images are applicable to be the dataset. The proposed system framework is developed based on Faster R-CNN with Inception-V2 architecture by fine tuning the training configuration parameters. The model is proposed to train up to 20,000 steps with the total loss is less than 0.07. The duration for the training process is about 64 hours. In terms of performance evaluation of the model, it is based on detection evaluation metrics applied by COCO. It shows that the mAP for Intersection Over Union threshold of 0.50 obtained 97.75% for the model in localizing the object along with the classes and the mAR with 10 detections per image obtained 72.03% for the model in classifying the object. In terms of overall detection accuracy, it obtained 88.71% for the accuracy. |
|---|