Skip to content

Latest commit

 

History

History
198 lines (138 loc) · 8.2 KB

readme.md

File metadata and controls

198 lines (138 loc) · 8.2 KB
Raw

Midterm - Human Object Detection

Build Status

Coverage Status

License: MIT

Authors and Roles

Phase1:

Driver- Aniruddh Balram

Navigator-Mayank Sharma

Design Keeper- Joshua Gomes

Phase2:

Driver- Joshua Gomes

Navigator-Aniruddh Balram

Design Keeper-Mayank Sharma

Introduction

Human (N>=1) obstacle detector and tracker based on C++ and OpenCV that employs a computer vision algorithm for location and categorization of humans(N>=1) in the picture. We are seeking to create this tracker utilizing a monocular camera, directly usable in a robot’s reference frame according to the need specifications supplied to us by ACME robotics.

For human recognition and tracking, we will use the robust YOLO neural network model trained on the COCO - 2017 dataset, which is one of the most accurate real-time object detection techniques.

Diagrams

UML Class Diagram:

alt text

UML Activity Diagram:

alt text

Tasks complete

  • IB: 1.101 Get preprocessing working
  • IB: 1.102 Get postprocessing working
  • IB: 1.103 Setup coverCV
  • IB: 1.105 Create an iteration development branch/ development branch
  • IB: 1.106 Select and add a software license as a file named LICENSE
  • IB: 1.107 Update UML
  • IB: 1.110 Update readme
  • IB: 1.112 Create classes for program
  • IB: 1.113 Implement cpplint and cppcheck
  • IB: 1.114 Create proper comments and revise old ones
  • IB: 1.115 Update Cmake
  • IB: 1.108 Create a docs directory with generated Doxygen files,
  • IB: 1.109 Create unit tests and test coverage,
  • IB: 1.111 URL of a 3 minute (max) video explaining the Phase 1 status of your API

Results:

The application currently performs well on single images. For demo purposes, we have used image from The Beatles album-art "The Abbey Road":

Input Frame:

WhatsApp Image 2022-10-19 at 9 26 58 PM(1)

Humans detected in the frame can be seen below:

Output Frame:

The application predicts the main 4 humans as well as the one behind (barely-visible)

Output

Task completed partially (Might contains errors due to build method):

  • IB: 1.104 Setup Github CI

Task incomplete:

Spreadsheet and Sprint Meeting Document Link

Spreadsheet link: https://docs.google.com/spreadsheets/d/1zVApmpAVnc7thu606UrYKJ7nqtlWpkH1Bv99EHn_2Is/edit?usp=sharing

Sprint Meeting Document Link: https://docs.google.com/document/d/154Ga8EMY9PfcyO2QEEYlObfffHEXi2clT9GDjFAgel4/edit?usp=sharing

Phase 1 Status video:

https://drive.google.com/file/d/1SrriQnXhLH50-QhWuF40HRacuLYLgIPe/view?usp=sharing

Known Issues/Bugs:

  • Error with the badge, it constructed but shows build:failed
[100%] Resetting code coverage counters to zero.
5Processing code coverage counters and generating report.
6Deleting all .da files in . and subdirectories
7Done.
8[==========] Running 1 test from 1 test case.
9[----------] Global test environment set-up.
10[----------] 1 test from dummy
11[ RUN      ] dummy.should_pass
12[       OK ] dummy.should_pass (0 ms)
13[----------] 1 test from dummy (0 ms total)
14
15[----------] Global test environment tear-down
16[==========] 1 test from 1 test case ran. (0 ms total)
17[  PASSED  ] 1 test.
18Capturing coverage data from .
19Found gcov version: 9.4.0
20Using intermediate gcov format
21Scanning . for .gcda files ...
22Found 3 data files in .
23Processing test/CMakeFiles/cpp-test.dir/main.cpp.gcda
24/home/runner/work/ENPM808X---Midterm-Project/ENPM808X---Midterm-Project/build/test/CMakeFiles/cpp-test.dir/main.cpp.gcno:version 'A75*', prefer 'A94*'
25geninfo: ERROR: GCOV failed for /home/runner/work/ENPM808X---Midterm-Project/ENPM808X---Midterm-Project/build/test/CMakeFiles/cpp-test.dir/main.cpp.gcda!
26make[3]: *** [CMakeFiles/code_coverage.dir/build.make:74: CMakeFiles/code_coverage] Error 255
27make[2]: *** [CMakeFiles/Makefile2:137: CMakeFiles/code_coverage.dir/all] Error 2
28make[1]: *** [CMakeFiles/Makefile2:144: CMakeFiles/code_coverage.dir/rule] Error 2
29make: *** [Makefile:169: code_coverage] Error 2
30Error: Process completed with exit code 2.
  • Two team members cannot build the program. Might have to do something with incorrect installation
  • Additional cpplint errors are mentioned in Results/cpplint.txt

Notes

  • Doxyfile is found in Code/doc_directory/Doxyfile

  • Tasks IB: 1.108 and IB: 1.104 may be implemented incorrectly due to build methods

Overview and purpose:

This program allows an image to be fed into the program, creating bounding boxes around every detected human. The goal is to use this program and make a human tracker out of it, with the input being a video feed.

In the current program, the input image is a picture of people on a zebra-crossing (found in ./app). The methods to build and run the program is shown in the bottom of this readme. The header file constants.hpp, defines important constant values for the program such as blob size, image size, filter thresholds, interface colours and font properties. The header file object_detection.hpp initializes the two classes used in this program. Both classes' methods are defined in the implementation file found in ./app.

The two classes used in this program are BlobGenerator and HumanObjectDetector.

The class BlobGenerator converts an image into blob and allows us to retrieve it.BlobGenerator::generateBlobFromImage() allows an image to be inputted and creates a blob from it.

BlobGenerator::getBlob() returns a blob as Mat datatype.

Derived from BlobGenerator by inheritance, the HumanObjectDetector class is used to detect humans in an image. This includes methods to draw bounding boxes, creating labels, prediction(pre-processing) and post-processing.

HumanObjectDetector:: labelBox() is used to draw a label around the 'class-text,' creating a label for the detected object, which is contained in a rectangle.

HumanObjectDetector::preProcessAlgorithm() forward propagates the blob into the yolo network. It is trained on COCO-2017 dataset to obtain properties such as confidence and class prediction.

HumanObjectDetector:: postProcessAlgorithm() filters out the low confidence and low-score classes matches labels the prediction with highest probability class.

HumanObjectDetector::applyNMSAndAppendRectanglesToImage() applies Non Maximal Supression and eliminates the redundant overlapping bounding boxes.

Dependencies:

  • opencv-4.x
  • eigen
  • lcov

To install dependencies:

# switch to repo directory
cd <directory_of_repo>

# run bash-script which holds all dependencies
bash dependencies.sh

Generate Doxygen document:

  • Step1: creates a Doxyfile
    Doxygen -g

  • Step 2: Edit the Doxyfile (INPUT and PROJECT_NAME)

  • Step 3: To generate html and latex folder
    doxygen ./Doxyfile. (Once thats done, two folders will be created html and latex, the html folder has index.html which will have the doxygen data)

  • Step 4: INPUT parameter in Doxyfile is the files you want to run doxygen on PROJECT_NAME parameter is the name of the title

Steps to Run test:

cd <directory_of_repo>/

# Create build directory and switch to it
mkdir build && cmake .. && make

# Run test file 'cpp-test' within test folder
./test/cpp-test

Steps to Build and Run Demo:

# Create build directory and switch into it
mkdir -p build && cd build

# Configure
cmake  ../opencv

#Build
cmake --build .

cd <directory_of_repo>
bash run_detector.sh