Prinsphield / ILFPT

Integrated Learning Framework for Pedestrian Tracking

Integrated Learning Framework for Pedestrian Tracking

By Taihong Xiao and Jinwen Ma

Introduction

This repo is an Matlab implementation of our paper. ILFPT is in integrated framework designed for pedestrian tracking, especially in the surveillance videos.

Here is a video demo.

Requirements

1. Caffe (see external/caffe/)
2. MATLAB
3. GPU: GTX 980/1080, Tesla K20/K40/K80

Preparation for Testing:

1. git clone --recursive https://github.com/Prinsphield/ILFPT.git
2. Build Caffe in the external/caffe
3. Download the test videos from either Google Drive or BaiduPan and extract them into the 'test/' directory
4. Download the detection network model from either Google Drive or BaiduPan and extract into models/ directory
5. Download our pretrained detection network from either Google Drive or BaiduPan and extract into output/ directory
6. Download the pretrained re-id network files from either Google Drive or BaiduPan and extract into reid_net/ directory

Testing Demo

1. Start Matlab from the root directory
2. Run faster_rcnn_build.m
3. Run startup.m
4. Run demo.m

Training Your Pedestrian Detection Network

1. Download SVD-B training data from OneDrive. Extract them into dataset/ directory and rename to VOCdevkit2007/
2. Modify related files in models/ dir to config detection network
3. Run scripts in experiments/ accordingly to train a detection network.

Note that the GPU cost for training a detection network is much higher than that for testing. Before training your own detection network, please ensure that your GPU memory memory meets the following requirement:

• 3GB GPU memory for ZF net
• 9GB GPU memory for VGG-16 net