pinyanLiu / computer_vision_1

homework 1 of computer vision

Computer_vision Homework

homework of computer vision

Pre-setting Up

Required Packages

• cmake or ccmake 3.10
• git
• opencv 4.4.0
• opencv_contrib 4.4.0

$ sudo apt install git
$ sudo apt install cmake
method of building opencv and opencv_contrib in linux,please refer to the follow tutorial link:
https://docs.opencv.org/4.4.0/d7/d9f/tutorial_linux_install.html
Please make sure to make install the opencv library

Build Method

$ git clone https://github.com/pinyanLiu/computer_vision_1
$ cd computer_vision_1/
$ mkdir build
$ cd build
$ cmake ..
$ make

execute method

with opencv

./build/util/with_lib 1.path to the video 2. level for example
./build/util/with_lib ./with_lib MOT17-09-SDP-raw.webm 5

Final Project describe

program describe

1. which is processing in build/util/with_lib
2. which gives you 6 pairs of result : video of each level and the answer txt of each level
3. Use CV::Tracker for object tracking
4. Use CSRT as the main tracker
5. When the object been blocked by another object,especially Level3 and Level4,stop the upgrade of cv::Tracker
6. restart upgrade after they have crossed by

---

Fourth Homework Describe

Template Matching

Program Describe

1. Which is processing in build/util/with_lib
2. Which gives you five result '30.png','60.png','90.png','120.png','150.png' locate at build/util , these are the results per 30 frame
3. If you use other video , it can also be executed
4. Capture frame per 30 frame and execute template matching
5. Generate Templates by decompose it into 40*40 with frame 'preFrame'
6. Using template matching method 'cv.TM_SQDIFF_NORMED'
7. Normalize the result between 0 ~ 255
8. Use cv::minMaxLoc() to get the miniest value point(which is the best matching point)
9. Paste the template on a new Mat on location we got from last step
10. Redo until there's no new frame be captured

---

Third homework describe

program describe

with opencv

1. which is processing in build/util/with_lib
2. which gives you four result 1.cannyBGR.bmp 2.bird.bmg 3.bird_crop.bmp 4.stitch.bmp locate at build/util
3. use canny() to detect edge
4. use HoughLinesP() to find line and colored with RGB
5. use getPerspectiveTransform() to get the transform matrix
6. use warpPerspective() to transform origin img to warping data
7. create stitcher to stitch two picture

without opencv

1. which is processing in build/test/warpWithCpp
2. which gives you two result 1.World.bmp 2.Image.bmg locate at build/test
3. lib is locate at lib/img_warp
4. a very fatal mistake i made is that i set the initial variables with 'int' which makes the whole transform incorrect,they should be set as 'float' or 'double' . It cost me two days for this stupid bug :(
5. imageToWorld() using X() and Y() to transform each pixel,and then forward warp the origin img
6. worldToImage() using U() and V() to transform each pixel,and then inverse warp the origin img

Second homework describe

program describe

with opencv

1. which is processing in build/util/with_lib
2. using opencv functions such as cvtcolor(),threshold(),dilate(),erode(),connectedComponentsWithStats()

without opencv

1. which is processing in build/test/no_lib
2. My binarizing method is different from the general one ,cause i didn't write the gray_scale function,so it may only usable in this case.
3. I use while loop check every pixel from top-left to bot-right,and then do it again in the inverse direction utill no changes happended.
4. I spend to much time on debuging the connectedcomponent part,so I didn't finish the remain part.

First homework

program describe

with opencv

1. which is processing in build/util/with_lib
2. using opencv functions like imread() imwrite() rotate() resize() to easily make changes to pictures.
3. Mat is a convenient type to read rgb data ,so getting r,g,b pic respectively,just set the other two channel to 0.

without opencv

1. which is processing in build/test/no_lib
2. I construct a class for this HW which contain the FILEs, the structure type of rgb,and three function for doing readwrite,channel_separation,and clock_wise_rotation.
3. In the constructor , First reading the path which user type in,and get the input image.Then it will open file for the rest of pic by calling fopen().
4. read_img() : cause we are not doing any change to the pic ,so i just use while() to read write the whole img data.
5. channel_separation() : separate two phase to read the img. First,read the header,which is the first 54 unsigned char.Second,using two for loop to read data in the rgb structure.Then,for each channel we want to sepatate,just need to set the other channels to zero and write the remain channel .
6. clock_wise_rotation() : cause the HW only request for a clock wise rotation, so I would rather not use a rotation matrix. Use two for loop, change the start point and the direction of writing img.