Berkeley Autolab

With the rise of live video streaming, a massive amount of data flows through the internet without getting collected. On the other hand, the recent advance in deep learning has shown the importance of big data. This repo features a pipeline for live stream video collection, object recognition in videos, and prepares the data to be comsumed by a 2D intersection traffic simulator.

numpy (1.14) OpenCV (>= 3.0) scipy (0.17.0) scikit-image (0.13) tensorflow (>= 1.0) youtube-dl (2018.2.4)

To use "SSD Detector", see SSDDetector.py.

To obtain pre-trained weights, download SSD-300 VGG-based from the SSD-Tensorflow repository. Afterwards, unzip VGG_VOC0712_SSD_300x300_ft_iter_120000.ckpt.zip, and place the contents (*.ckpt.index and *.ckpt.data-00000-of-00001) under SSD/checkpoints.