Skip to content

roboskel/hpr

Repository files navigation

#Installation Instructions: Download this ROS package inside a catkin workspace.

#Working with the below library versions:

scikit-image==0.12.3 scikit-learn==0.15.1 scipy==0.13.3 numpy==1.10.1 matplotlib==1.3.1

It is really important to install the indicated versions of the libraries because older versions are much slower. (pip recommended)

RECOMMENDATIONS:

Use same timewindow, and wall set time for each training set, and use the same values when
evaluating.

#Running out-of-the-box roslaunch human_pattern_recognition hpr.launch The script is now waiting for laser scans to be published in the topic /scan. If you want to change the parameters of the script, just edit hpr.launch, or create your own .launch file.

Human-Pattern-Recognition

Real time recongition of humans through laser scans

#Python files explained #----------------------

#I)bag2mat.py:

Convert a .bag file to .mat format.

Command line use:

python bag2mat.py <.bag_file_path> <.mat_file_path> <laser_scan_rostopic> <scan_duration>

#II)annotate.py :

Annotate (label) each cluster as human or not human, in order to train the classifier later. annotate.py generates many .p files, some of which are classifiers trained on each and only file annotated. This means that every time the annotation process ends, a new classifier is created trained only on the .mat that was just annotated.

Command line use:

python annotate.py <time_window> <number_of_frames_to_create_walls> <.mat_file_path>

#III)merge_train.py:

This is an old file. Use recognition_procedure instead. Merge_train uses all the annotations from a folder, to create a classifier based on all of those annotations.

Command line use:

python merge_train <folder of annotated .mat files>

#IV)hpr.py:

Runs the human pattern recognition classifier.

Command line use:

roslaunch human_pattern_recognition hpr.launch

#V)generate_metrics.py

Runs the mat file that the classifier was tested on, and lets the user annotate the same clusters as the classifier in order to generate some basic metrics (Precision, Recall, Accuracy).

Command line use :

python generate_metrics.py </path/to/classification_results.mat>

#VI)recognition_procedure.py The python script that is now used to train the classifier. Three different classifiers can be trained, but LDA seems to be the best.

#VII)Library files gridfit.py, myhog.py and mytools.py are libraries used to run the code. They were pre-written by others or ported to python.

About

Recognition and tracking of human walking in laser scans

Resources

Stars

Watchers

Forks

Packages

No packages published