Autonomous Driving Perception - Semantic Segmentation and Object Detection

This repository is for Computer Vision course project at George Washington University, The main purpose of this project is trying to complete Autonomous Driving Perception tasks to ensure safe driving condition.

Team Member

• Zhangmingyang Su, Ze Gong, Derasari Preet.

Package Installation

• Tensorflow
• OpenCV
• yolov3.cfg
• coco.names
• yolov3.weights(Link because of the large file)

Semantic Segmentation

DeepLab Model Architecture

Segmentation Result

After Implementing DeepLab model, The driving scene is segmented into 16 categories which represented by different colors in real-time.

Segmentation result based on different frame rates

From the graph, it's easy to figure out the higher frame rate is, the better performance is. There are lots of things happen in a short-time, in order to ensure safe driving condition, we need to capture more information from frame to frame.

Segmentation result of original data vs temporal Data

For the temporal data, such as the transformation of the color, background, distortion, etc. All of these frames are not independent, so, we sum the current frame result and previous frame result to smooth the prediction. By using the temporal data, the performance is a little bit better than the original one without the temporal data.

Object Detection

YOLO-V3

Object Detection Result

Usage

1. Download dataset from MIT-DriveSeg.
2. Run object_detection.py to implement yolov3 model.
3. Run data_evaluate_segmentation.py to see performance change based on different frame rate(manually).
4. For orginal data vs temporal data analysis, Run data_evaluate_segmentation.py to see performance change in terms of Class IOU Metric.
5. More information, check course paper.