Team: Primage Imocessors

Project for Digital Image Processing (CS7.404.M21) - Monsoon 2021

Project Representatives:

1. Aditya Kumar Singh - 2021701010
2. Bhoomendra Singh Sisodiya - 2021701037
3. Dhruv Srivastava - 2021701021
4. Prateek Jaiswal - 2021701009

Quick links

1. Overview
2. Directory Structure
3. Instructions
4. Results
5. Dependencies

Overview

The aim is to design a vehicle detection and counter using a combination of different video-image processing methods including object detection, edge detection, frame differentiation, and Kalman filtering.
The presented code includes implementation with two different approaches-

• Using previous frames for motion detection as proposed in link
• Constructing a background using moving-averaging and using it to detect motion from current frames.

Using automated methods, one can execute better traffic management methods, such as changing the timings of traffic lights based on traffic flow, traffic surveillance, etc. We’ll mainly focus on vehicle detection on roadways and classify the passing vehicles in different specific types. The video data used for this project can be downloaded from here.

Flow chart of proposed improved method

Paper Method	Improved Method

Directory Structure

├── data
    ├── 1.mp4
    ├── 2.mp4
    ├── 3.mp4
    ├── 4.mp4
    └── 5.mp4
├── images
    ├── 1.gif
    ├── 2.gif
    ├── 3.gif
    ├── 4.gif
    ├── 5.gif
    ├── bg_creation.gif
    ├── flow_chart.gif
    ├── kalman_filter.gif
    ├── prev_prev_frame_issue.png
    ├── region_selection.gif
    ├── single_lane_detection.gif
    └── sobel_edge_det_issue.png
├── KalmanFilter.py
├── gui.py
├── project_ppt.pptx
├── requirements.txt
├── runner.py
├── vehicle_detection.py
└── README.md

Instructions to run the project and replicate the results

1. Create a virtual environment

$ virtualenv -p python3 venv

If virtualenv is not installed on your system, it is recommended to use pip for the same.
($ pip install virtualenv)

2. Activate the virtualenv and install the project requirements

For Ubuntu/MacOS-

$ source venv/bin/activate
$ pip install -r requirements.txt

For Windows OS-

venv\Scripts\activate
pip install -r requirements.txt

3. Execute the GUI script

$ python gui.py

4. Instructions to operate the GUI.py.

1. In Clip Selection Menu choose one clip on which you want to perform vehicle tracking. (Clip x is a short video from a traffic camera, where each video is different from other.)
2. Create Dynamic Background option: Allows one to manually select the ROI (Region of Interest) or detection zone by dragging the mouse followed by pressing q for confirming the region selected.
   
   
3. Paper Replicate option:
   1. If True then one will be reproducing the results from paper.
      1. Video frames Preprocessing:
         1. Gray Scale conversion of original RGB frames.
         2. Gamma Transformation of gray frames with gamma factor = 1.2.
         3. Applying Sobel filter to detect edges followed by thresholding to convert it into binary images.
      2. Motion Analysis:
         1. To detect vehicle, background subtraction is done by frame differentiation method where commonality of Nth and (N-1)th binary frames is subtracted from the commonality of (N+1)th and Nth binary frames.
            
            
         2. Next, a detection zone is defined manually as mentioned above (on which all operations are carried out).
         3. For each moving vehicle a KalmanFilter object is assigned (which also acts as it's ID) to track and estimate it's position (or co-ordinate) as well as minimize noise disorders. Although edge detection can find moving objects, the Kalman filter makes an optimal estimate of positions based on a sequence of localization measurement.
         4. Now the total counts of such ID (which is a list consisting of vehicle's position info, KalmanFilter object, and its bounding box area) will not only give us the total count of vehicles passed but also which type of vehicles have passed using bounding box's area information.
   2. Else, we'll be using our own Modified Method.
      1. Constructed a 'static' backgorund using Moving Average method.
         
         
      2. Video frames Preprocessing: Applied GrayScale conversion followed by GaussianFiltering to smoothen out the artefacts and unnecessary noises.
      3. Execute background subtraction by taking absolute difference (cv2.absdiff()) between current video frame and static background.
      4. Thresholding followed by Dilation ensures that the (disconnected components of) moving vehicle in obtained binary image comes out as a single "white blob" (or component).
      5. Using nearest neighbor criteria, connect a given vehicle's position in each frame to track its bounding box so as to maintain the counter.
   3. A total of four classes are there on the basis of size of vehicle, namely,:
      • Type1: bicycles, motorcycles.
      • Type2: motorcars.
      • Type3: pickups, minibuses.
      • Type4: buses, trucks, trailers.
   4. Using bounding boxes area, we try to predict the class of vehicle and update their class counter.

Our Results

For modified (our) method.

DETECTION TEST RESULTS

Actual vehicle passed	Counted number of vehicles by the method	Number of errors
44	42	4.54%

Confusion Matrix

CF Matrix	Type 1	Type 2	Type 3	Type 4
Type 1	0	0	0	0
Type 2	1	28	0	0
Type 3	0	0	10	0
Type 4	0	0	0	3
Noises	0	0	0	0

For method mentioned in paper.

DETECTION TEST RESULTS

Actual vehicle passed	Counted number of vehicles by the method	Number of errors
44	58	18.18%

Confusion Matrix

CF Matrix	Type 1	Type 2	Type 3	Type 4
Type 1	1	0	0	0
Type 2	2	25	0	0
Type 3	0	1	10	0
Type 4	0	0	2	3
Noises	8	0	0	0

Dependencies

cycler==0.11.0
kiwisolver==1.3.2
matplotlib==3.4.3
numpy==1.21.4
opencv-python==4.5.4.58
Pillow==8.4.0
pyparsing==3.0.6
python-dateutil==2.8.2
six==1.16.0
tk==0.1.0