CBCL_FasterRCNN_MATLAB

Module to train a faster RCNN using the MIT CBCL dataset for vehicle detection, using MATLAB

CBCL dataset here: cbcl.mit.edu/software-datasets/streetscenes/

build_data() may be used standalone to extract bounding box annotation information from the CBCL annotations. CBCL annotations in XML format have more than 4 points for each annotated vehicle, build_data() converts that information to bounding box information that can be used with a faster RCNN object in MATLAB.

This:

Becomes this:

Use 'Mod_MIT_training.m' to train your detector, change parameters in the options variable. 2 pre-trained detectors in the "detectors" folder, 1 trained on 300 random images and another on 900 random images from the CBCL dataset. (These detectors have horrible accuracy, you have been warned!)

detection_with_faster_r_cnn.m uses the faster RCNN to detect vehicles from video detection_bgsub_faster_rcnn.m applies a gaussian mixture model to extract only moving objects, then uses the trained faster RCNN to detect if those moving objects are vehicles or not. This performs faster than just a plain faster RCNN, accuracy is reduced though.

TODO:

• Video links

• Organize code

• Write a better readme