Adaptive Computer Vision Algorithm for Real Time Moving Vehicle Detection and Segmentation

Project: This project describes an adaptive learning approach to detect, segment, measure and track objects in outdoors when there is no training data by applying computation geometry, topology and engineering physics.

Problem

No training data for object detection
Assumption: Stationary camera at traffic signals and buildings

Solution

Adaptive Object Detection

• Involves frame subtraction to separate background and foreground
• background and foreground detection using Gaussian Mixture Model
• Applying morphological operations and filters to remove noise
• Implementing Canny Edge detection for corner and boundary of the moving object
• Generating persistence graphs and barcodes to store object positions

Polygon Object Instance Segmentation

• Creating a feature vector to store boundary points and centroid of moving object
• Triangulating using boundary points and centroid to implement polygon segmentation on moving object
• Generating graph using boundary points and centroid to store the previous positions of moving object

Shape Analysis for Object Classification and Recognition

• Generating the feature vector of the moving object
• implementing meta learning for one shot learning for object classification
• comparing results with Siamese Neural Network

Object Tracking

• Applying kalman filter for object tracking