openmht

Python module for multiple hypothesis tracking. Based on the article:

C. Kim, F. Li, A. Ciptadi and J. M. Rehg, "Multiple Hypothesis Tracking Revisited," 2015 IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 2015, pp. 4696-4704, doi: 10.1109/ICCV.2015.533.

URL: https://ieeexplore.ieee.org/document/7410890

This implementation utilizes motion scoring only (no appearance scoring)

Installation

Install the latest version of Python 3

$ pip install openmht

To also plot tracks after completion, install matplotlib:

$ pip install matplotlib

Formatting the Input CSV File

Format the input CSV columns with frame number and pixel positions using the examples under SampleData/ as a reference. The U,V values represent the 2D positions of objects/detections in that frame. A value of None in the output CSV indicates a missed detection. The Track column indicates the final track ID for a detection.

MHT Parameters

Modify parameters by editing the params.txt input file. Please read the paper mentioned above to understand how these parameters can be updated to improve performance and accuracy:

Motion scoring parameters

Parameter	Description
v	The image (frame) area in pixels (Default: 307200).
dth	Gating area for new detections implemented as the threshold for the Mahalinobis distance d2 between the observation and prediction (Default=1000).

Kalman filter parameters

Parameter	Description
k	Gain or blending factor. Higher gain results in a greater influence of the measurement relative to the filter's prediction (Default=0).
q	Initial estimate of the process noise covariance (Default=0.00001).
r	Initial estimate of the measurement noise covariance (Default=0.01).
pd	Probability of detection PD (Default=0.9).

Track tree pruning parameters

Parameter	Description
n	Look back N frames and prune all branches that diverge from the solution. A larger N yields a more accurate solution due to a larger window, but will take a longer time (Default=1).
bth	If the number of branches exceeds the number Bth then prune the track tree to only retain the top Bth branches.
nmiss	A track hypothesis is deleted if it reaches Nmiss consecutive frames of missing observations.

Running the Program

OpenMHT takes in the input CSV detections and the parameter file, and saves to the provided output CSV file:

$ python -m openmht InputDetections.csv OutputDetections.csv ParameterFile.txt

A default parameter file is provided in this repository: params.txt

For generating track plots, add the --plot parameter (requires matplotlib):

$ python -m openmht ... --plot

Example Results

Results from running SampleData/SampleInput.csv:

Plot

Data

Input

Output

Frame U V 0 0.0703 0.3163 1 0.1071 0.3746 1 0.1325 0.1618 2 0.1694 0.4534 2 0.1809 0.1910 2 0.4205 0.0977 3 0.2200 0.5700 3 0.2408 0.2755 3 0.5081 0.1618 4 0.2938 0.6429 4 0.3007 0.3222 4 0.5703 0.2201 5 0.3445 0.7157 5 0.3767 0.4184 5 0.6555 0.2988 6 0.4297 0.8149 6 0.4459 0.4767 6 0.7247 0.3688 7 0.4850 0.8703

Frame Track U V 0 0 0.0703 0.3163 0 1 None None 0 2 None None 0 3 None None 1 0 0.1071 0.3746 1 1 0.1325 0.1618 1 2 None None 1 3 None None 2 0 0.1694 0.4534 2 1 0.1809 0.191 2 2 0.4205 0.0977 2 3 None None 3 0 0.22 0.57 3 1 0.2408 0.2755 3 2 0.5081 0.1618 3 3 None None 4 0 0.2938 0.6429 4 1 0.3007 0.3222 4 2 0.5703 0.2201 4 3 None None 5 0 0.3445 0.7157 5 1 0.3767 0.4184 5 2 0.6555 0.2988 5 3 None None 6 0 None None 6 1 0.4459 0.4767 6 2 0.7247 0.3688 6 3 0.4297 0.8149 7 0 None None 7 1 None None 7 2 None None 7 3 0.485 0.8703