Video Vehicle Detection and Classification system
Rahul Wadbude, Shubham Agrawal, Avani Samadariya, Ritika Mulagalapalli
The Crux
Given an input traffic video, the task was to detect moving objects and then classify them into vehicles, pedestrians etc.
Used MOG/MOG2 algorithms from python opencv for extracting foreground objects from videos.
Individual objects from images were cropped and SIFT/deepnet features were extracted.
Extracted features were finally fed to classifier for classifying into car, bicycle, auto etc.
Results and Analysis
Classification Accuracies
The following are the results obtained for both the features using one-vs-rest svm for comparison:- SURF features with tuned parameters : Kernel = Sigmoid, C = 10000; gave an accuracy of 59.94%.
- Deepnet features with tuned parameters : Kernel = Polynomial, C = 1; gave an accuracy of 94.4297%.
Prediction on video input :
