CV in Detection & Tracking | Prateek Bhargava

Computer Vision (CV) in Detection and Tracking

Arrow Detection and Tracking in Real Time

~~Coded In Python, C++

In this project, object tracking algorithms are built to identify the presence and orientation of a green arrow in the four directions: Up, Down, Left, Right. This operation is performed in real-time using raspberry pi and picamera.

Traffic Light Detection in Real Time

~~Coded in Python

In this project, object tracking with the Raspberry Pi and OpenCV to simulate a camera used for navigation onboard an autonomous vehicle is implemented. Here the green light in a traffic light is detected and tracked, where the steps taken to achieve this are as follows:

Step 1) HSV masking

Step 2) Finding contours and drawing bounding circles

Step 3) Applying it to picamera feed for real-time detection and tracking

Lukas Kanade Tracker

~~Coded in Python

Implementation of Lucas Kanade tracker on a set of three given video sequences: featuring a car on the road, a baby fighting a dragon, and Usain Bolt’s race respectively. The Lucas Kanade tracker is set to find the optical flow of events inside a frame which is initialized with the help of a rectangle bounding the objects to be tracked. The objects to be tracked are collected in a frame with the help of the bounded and adjacent pixels. For each frame, the intensity between the pixels of the two frames are compared to find the vector flow or the flow of motion of the objects. Finally, the optical flow equations are used for the pixels and the centered window.

Color Segmentation using GMM

~~Coded In Python, C++

The aim of this project was to train a learning model to identify buoys of different color distributions present in the underwater video sequence. The video sequence shows almost circular buoys of three different distinct colors, namely, yellow, orange, and green. Each buoy present in the video is to be detected by bounding contours of their respective colors. Conventional segmentation techniques involving color thresholding do not work well in environments such as underwater. The concept used to achieve our objective is color segmentation using Gaussian Mixture Models and Expectation Maximization techniques. In this method, the color distributions of the buoys are ”learned” from the data-set extracted from the video sequence, and then the ”learned” model is used to segment them.

Enhancing Low Light Frames

~~Coded in Python

Enhanced low-light image frames by 30% via application of CLAHE & gamma correction techniques. In this method, the image is divided into blocks of 8x8. where then a histogram is plotted to check whether to clip the blocks or not. Then CDF is computed and transformation function for each block in the center pixels. Then the remaining pixels are transformed with respect to the center pixel. Now we use the convolution method to apply the filter and then use the CLAHE method. In this, we place our filter in one corner of the image, such that it overlaps with 9 pixels present in the corner

Before

After

Turn Prediction & Lane Detection

~~Coded in Python

The concepts applied in this project for Turn Prediction and Lane Detection Warp-Perspective, Homography, and Histogram Equalization.

Augmented Reality (AR) Tag Detection

~~Coded In Python

This project focuses on the applications of Augmented Reality where custom-built tags are detected in a frame of video. The project is divided into two phases. The first phase involves the detection and tracking of the tag using contour and corner detection. The second phase involves the image processing where the detected tag is first replaced or superimposed by a provided template image (Lena.png) and then a virtual 3D cube is placed over the tag.