Student Projects

VE/VM450

Vision Detection for Quality of Automotive Tail Lamp

Instructors: Prof. Jigang Wu
Team Members: Dooho Ro, Fenglei Gu, Sarit Kittirattanapaiboon, Yuankai Zhou, Xiheng Wang (in alphabetical order)

Project Video

Team Members

Team Members:

Dooho Ro, Fenglei Gu, Sarit Kittirattanapaiboon, Yuankai Zhou, Xiheng Wang
(in alphabetical order)

Instructors:

Prof. Jigang Wu

Project Description

Problem

Tail lamp is a key component of automotive lighting system. During manufacturing, traditional approach of tail lamp quality detection with human eyes is slow, error-prone and expensive. Using new method of detection based on Computer Vision, this project aims to improve the detection speed and accuracy, which can which can greatly improve the efficiency of the factory.
Fig. 1 Image of an automotive tail lamp

Concept Generation

In order to complete this project, there are two steps to be considered.

The first step is image processing which extracts relevant features from source images of tail lamps.The second step is classification, where the classifier algorithm decides whether the lamp is of good or bad quality, based on the selected features from the previous step.Finally, we also need to consider which programming language to use for implementation.

Fig. 2 Overall concept diagram
Fig. 3 Brainstorming diagram

Design Description

Image Processing with SED:
As seen in Fig. 2, the process starts from static images taken from a video of a tail lamp. These images are then analyzed through the Sobel Edge Detection method to extract the features. The result from the method is seen in Fig. 4
Fig. 4 Analyzed image of a tail lamp
Classification with SVM:
These features are then put into a SVM classifier in a binary sequence. To have an efficient SVM on a certain pattern, corresponding features and those from different patterns are introduced as comparison and data augmentation in classifier training. As seen in Fig. 5, SVM makes a reliable classification with maximized margin, to decide if the input features fits the standard of a certain pattern.
Fig. 5 Support vector machine [3]
Programming with Python 3:
There are abundant open-source libraries available for use, such as sklearn and OpenCV, which will significantly helps the development of the project. In the meantime, its runtime performance satisfies the requirement of the factory.

Validation

Validation Process – Cross Validation:

Measures of Performance:

  • Accuracy: >80%
  • Precision: >0.99
  • F1 Score: >0.8
  • AUC: >0.8
· To be further determined and subject to change. Right of final interpretation reserved.
Fig. 6 K-fold cross validation [1]
Fig. 7 Confusion matrix[2]

Conclusion

Overall, the goal of the project is to deliver an algorithm that can assist with the manufacturing of tail lamps. Current defective lamp detection methods use human eyes, which can be slow, erroneous, and expensive. For the sake of reliability and ease of control, we have chosen the methods that are the most straight-forward while keeping the accuracy as high as possible.

Acknowledgement

Special thanks to Hasco Vision and our instructor, Prof. Jigang Wu.

Reference

[1] https://en.wikipedia.org/wiki/Cross-validation_(statistics)
[2] https://en.wikipedia.org/wiki/Confusion_matrix
[3] https://www.codershood.info/2019/01/10/support-vector-machine-machine-learning-algorithm-with-example-and-code/

UM-SJTU JOINT INSTITUTE