Detect pedestrians in infrared by Haar Features Cascades. As a ground truth we use tracking manual marked subjects.

Dependencies:

pip install opencv-python

All configurable variables contain ./config.py file.

Runing:

py .\run.py

Every time, when:
-new sequence(folder) is loading,
-tracker lost object,
-space key was pressed,
Starts window for select objects for trucking. You need to mark object for frucking using your mouse, then press space, to confirm selection. After that you can select another object (pressing space one more time), or configm all selected objects by pressing return/enter.

For clear tracking object You should hit backspace in tracking object selection. Unfortunately old tracking boxes still be present on this window.

In real time you can see frame/image, with:
-green border box - tracked elements,
-blue border box - detected objects.

Inside folder ./images/ we assume exist several folders for different images sequences.

All image datasets is located into images.7z archive. The archive you can download from: https://cvlab.zut.edu.pl/pedestrian_detector/images.7z

The scripts for each sequences (folders in ./images/) save a .txt file inside ./results/ folder. Moreover, during computing display results in console.

Results format is:

1 1 0 0.546903

It means, for specific frame/image (in .txt file there is also image name/number):
-first number describes how many object was tracked (ground truth)
-second number describes how many object was properly detected
-third number describes wrong detections
-last number shows mean IOU for each detected objects on frame,

The Proper detected is, when IOU for detect and ground truth is equal or more than MINIMUM_IOU setting in ./config.py.

Function addPedestriansToTrack allow to mark multiply object for tracking.

Function calculateIOU calculateIOU for all detection, and count how many detections was correct.

Function detectPedestrianOnImage set tracking objects, detect pedestrians and returns IOU between them.

Function drawBoxes draw detect/tracking box on image/frame.

Function forEachImage explore ./image/ folder and each subfolder, and on every image runs passed function and collect results.

Function saveResults save computing results for each subfolder inside ./image/ to given file in ./results folder.