This is a undergraduate thesis project.
Topic: 3D reconstruction and object detection
This is work is mainly based on Shuran Song & Jianxiong Xiao's work and some improvement is made.
The original paper is on http://slidingshapes.cs.princeton.edu/
-
Using images captured from RGB-D camera (Intel RealSense) and using the information to build a 3D point cloud based on 2D images and depth information.
-
Using Solidworks to build an ideal model as template. The template then is used to compute the differences between input images and template.
-
Since the ideal template is CG model, a sliding window is introduced to detect target in the input images. The both the depth and height information is considered.
-
Pre_process uses a sliding window to do a rough search in the whole image. If there is an obvious similarity between input and template, then the target will be identified by a red box.
-
If the pre_process cannot give a result with high confidence, then the input image will be sent to the whole model which computes all depth features.
-
The depth features that the model computes include: TSDF, covariance matrix, normal vector.
-
A SVM is trained in the decision making process.