Combining Road Segmentation and Traffic Object Detection

To achieve better understanding of autonomous vehicle’s surroundings, we proposed a new pipeline by combining custom variants of SegNet for road segmentation and YOLOv1 for traffic object detection. By utilizing transfer learning and novel image augmentation that is not mentioned in the original configuration, our model obtained better performance on chosen datasets. We were able to achieve good results on both in-domain and out-domain traffic datasets captured in real life.

Writeup

For our proposal see HERE

For our project report see HERE

Model checkpoints

• SegNet: https://drive.google.com/drive/u/0/folders/1M_jH04mBN9ZuXHPCkZbCAFIRHUbnOhO7

• YOLOv1: https://drive.google.com/file/d/1SDcfLAAhujRvOCjaW9-325Ol7Lp9Vttu/view?usp=sharing

Visual results

On Camvid valid dataset:

On Camvid training dataset (Note that we used this dataset in our training):

On KITTI 1:

On KITTI 2:

Contact

{Bingzhao Shan, Songlin Liu, Zihan Wang, Zuoyi Li} @ Umich

Some intermediate SegNet result can be found here