Kernelized Correlation Filters GPU

A correlation filters and deep learning based object tracking algorithm. The algorithm is implemented on the GPU using CUDA C++.

Requirements

• ROS Indigo or higher
• CUDA (7.5 or higher)
• OpenCV 3.0 or higher
• image_view
• Caffe
• C++11

Downloading the Package

$ git clone https://github.com/iKrishneel/kernelized_correlation_filters_gpu.git
$ cd kernelized_correlation_filters_gpu
$ git pull --all

Compilation

The package is a ROS node and can be complied using the catkin tools

  $ catkin build kernelized_correlation_filters_gpu

If you dont use ROS, you can comment out the ROS dependencies in the following three files:

• CMakeLists.txt
• src/kernelized_correlation_filters_node.cpp
• include/kernelized_correlation_filters_gpu/kernelized_correlation_filters_node.h

Make sure to set other appropriate paths.

Running

 $ roslaunch kernelized_correlation_filters_gpu kernelized_correlation_filters_gpu.launch <options>

<options>

In the launch file there are number of options that can be set according to how you what the algorithm to run. Brief descriptions are as follows:

• image [sensor_msgs/Image]: - input image topic name
• headless [Bool]: - to visualize the intermediate results set it to false
• downside [Int]: - (>= 1) specifies the factor for image reduction
• init_from_detector [Bool]: - flag for alterning how the tracker will be initialized.
• runtype_without_uav [Bool]: - flag to specify if tracking on uav so the tracker can use odom for height based scale estimation

Results

The video results of this algorithm can be viwed here

Reference

Please cite our work should you find our algorithm useful.

Krishneel Chaudhary, Moju Chou, Fan Shi, Xiangyu Chen, Kei Okada, Masayuki Inaba: Robust Real-time Visual Tracking Using Dual-Frame Deep Comparison Network Integrated with Correlation Filters, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2017), Vancouver, Canada. September 24 - 28, 2017