Fast human detection in depth images

An implementation of Fast Human Detection for Indoor Mobile Robots Using Depth Images for Kinect V2 depth images.

Differences from the paper:

• Kinect V2 over V1
• Region planarity checks are disabled by default
• A neural net is used for classification instead of SVM
• Stratified sampling is used over random sampling for point cloud construction

Building

Standard CMake build.

Generate the Makefiles:

$ mdkr build
$ cd build
$ cmake ..

Or MSVC project files:

$ mdkr build
$ cd build
$ cmake -G "Visual Studio 12 Win64" -DCMAKE_PREFIX_PATH=KINECT_SDK_DIR ..

KINECT_SDK_DIR is usually C:\Program Files\Microsoft SDKs\Kinect\v2.0_1409

Run make or build the MSVC projects.

Training a dataset

fhd_ui can be used to create a training set from depth images.

"open database" selects a Sqlite DB of depth images. Clicking on a candidate (marking it green) sets it as a positive candidate (human). Pressing space commits the candidates to the database, where selected (green) marks a human and unselected candidate marks a negative candidate. X advances to the next frame

After creating the training set, a classifier can be trained under the Training tab.