Trajectory Post-Processing for Traffic Multi-Object Tracking

A lightweight toolkit for improving traffic multi-object tracking (MOT) outputs by:

• breaking long trajectories into shorter, reliable tracklets when motion anomalies occur
• linking tracklets back into complete object trajectories using learned pairwise matching
• smoothing and stitching trajectories for cleaner final outputs

Features

• TrajectoryBreakPhase: detects unstable trajectory points with a 2D Kalman filter and Mahalanobis distance
• KalmanFilter2D: constant-acceleration model for prediction and anomaly scoring
• LinkingPhase: matches tracklets using a logistic regression model and geometric/motion features
• Track: lightweight tracklet container for frame intervals and detection records

Requirements

• Python 3.8+
• opencv-python
• numpy
• onnxruntime-gpu
• pandas
• scikit-learn
• scipy
• supervision

The repository includes a minimal requirements.txt, but the code also depends on pandas, scikit-learn, and scipy.

Installation

Install dependencies with pip:

pip install -r requirements.txt
pip install pandas scikit-learn scipy supervision

Usage

Example pipeline using the main processing classes:

from trajectory_improvement.postprocessing import TrajectoryBreakPhase, LinkingPhase
from trajectory_improvement.tracklet import Track
from sklearn.linear_model import LogisticRegression
import pickle

# 1. Split long object trajectories into shorter tracklets
break_phase = TrajectoryBreakPhase(input_csv_filename='input_mot.csv', video_fps=30)
tracklets = break_phase.create_trackelts()

# 2. Load a trained logistic regression model for linking
with open('linking_model.pkl', 'rb') as f:
    log_reg_model = pickle.load(f)

# 3. Reconnect broken tracklets and save the final MOT output
link_phase = LinkingPhase(
    log_reg_model=log_reg_model,
    track_list=tracklets,
    csv_filename='output_mot.csv',
    input_csv_filename='input_mot.csv'
)
link_phase.run_post_process()

Expected Input Format

The input MOT CSV should contain at least the following columns:

• frame_number
• tracker_id
• x_center
• y_center
• bb_left
• bb_top
• bb_width
• bb_height
• class_name
• confidence

The LinkingPhase also relies on class labels and confidence scores to merge tracklets.

Module Summary

• trajectory_improvement.postprocessing
  • TrajectoryBreakPhase: processes each object trajectory, computes Mahalanobis distance per frame, and splits trajectories at anomalies.
  • KalmanFilter2D: performs prediction and update steps for 2D motion and returns innovation-based anomaly scores.
  • LinkingPhase: uses pairwise features and a logistic regression classifier to link tracklets into complete trajectories and exports the final CSV.
• trajectory_improvement.tracklet
  • Track: container object for a tracklet with start/end frames and detection sequence.

Notes

• The current implementation uses a constant acceleration motion model and may require tuning of the mahalanobis_distance_thresh, lost_track_tresh, and positive_match_thresh parameters for optimal results on different datasets.
• The LinkingPhase assumes that tracklets with very short lengths should be finalized rather than held as linking candidates.