Home

Awesome

MTMC

A paper list of Multi Target Multi Camera (MTMC) tracking and related topics <br/> including application case in: vehicle tracking :red_car: , pedestrian tracking :frowning_person: , sports player tracking :soccer: .

<details><summary>Click to show menu</summary> <p>
  1. <a href="#multi-target-single-camera-tracking-paper">Multi Target Single Camera Tracking Paper </a> <br/>
  2. <a href="#multi-target-multi-camera-tracking-paper">Multi Target Multi Camera Tracking Paper </a> <br/>
  3. <a href="#related-github-repo">Related Github Repo</a> <br/>
  4. <a href="#related-competition">Related Competition</a> <br/>
<!-- 5. <a href="#related-group-or-researcher">Related Group or Researcher</a> --> </p> </details>

Multi Target Single Camera Tracking Paper

2022

interesting to see a variant of SORT (observation-centered) achieve decent results

not tracking but seems applicable in MC-tracking, detect bbox from images and match roughly, use interesting GNN formulation to refine camera pose: image as node, edge as relative pose, bbox info added during message passing

2021

at first associate box with high detection score, then associate box with low detection score, improve tracking on occluded objects

instance similarity learning based on region proposal, flexible, no external data required

Transformer, detection and tracking simultaneously

2020

Deep Hungarian Net, approximate MOTA, MOTP for loss function directly

apperance embedding (node) and geometry distance embedding (edge) for graph, edge classification with cross entropy loss

pipeline: detection, feature extraction, affinity, association

end-to-end MOT, use adjacent frames (chained) to combine detection, feature extraction and tracking

2019

use appearance, location and topology cues for similarity score, then graph solved by Hungarian algorithm

GNN, Siamese network

motion and appearance extention -> Tracktor++

traditional and deep visual trackers

correlation filter, deep learning and convolutional features

2018

use epipolar geometry, tracklet as node in graph

online MOT tracker

2017

learn statistics to normalize effect of camera poses, temporal adjacent constraint for data association

not use appearance feature, very fast, not accurate

IoU tracker, no visual cues used, fast

RNN as tracker, LSTM for data association

2016

use Siamese CNN to learn similarity, for data association, graph solved by Linear Programming

2014

interaction between objects, relax the dependency of tracking on detections

Multi Target Multi Camera Tracking Paper

2022

step 1: single camera tracking & generate appearance feature, step 2: multi camera association with GNN (single camera trajectories as node, averaged feature as node feature, cos(feature) as edge feature), weighted loss for imbalance

2021

tracklet as node, link prediction for data association, ok for w/wo overalaping view, use large training data

detection-> feature extraction, homography -> cross-camera cluster -> incremental temporal association, small latency, not very accurate

2020

fusion all views into ground-plane occupancy heatmap

tracklet representation with spatial-temporal attention, then tracklet-to-target assignment

tracklet-to-target assignment

single camera tracklet -> multi-camera tracklet fusion with appearance and physical features

use TrackletNet for single camera trajectory -> inter-camera tracking

single camera tracking -> match tracklets across camera views

Reinforcement learning, collaborative multi-camera

camera synchronization, SfM, Bundle Adjustment, spline representation for drone trajectory

combine appearance and homography for hierachical clustering, known camera pose

2019

Centralized (combine cross-camera views before tracking, like Wen et al.) and Distributed methods (single-camera tracking before fusion)

single camera detection -> create/match to track, with apperance, motion, spatial-temporal cues (cross-camera)

2018

tracklet -> single camera trajectory (correlation clustering) -> multi camera trajectory

single camera tracking -> CNN feature extraction -> multi camera tracking (KMeans)

2017

3D position for affinity computation, need know camera parameters, cross-view coupling before trajectory

2014

two tracker (detection and regression) in parallel, measure their correspondence

2013

detection as node in hypergraph to find 3d reconstruction, which is node in a min-cost flow graph, solved by binary linear programming

2012

Related Github Repo

Related Dataset

Related Competition

<!-- ## Related Group or Researcher - [Dynamic Vision and Learning Group, TUM](https://dvl.in.tum.de/research/) > TrackFormer, Tracktor++, Siamese - [CVLab, EPFL](https://www.epfl.ch/labs/cvlab/research/research-surv/research-body-surv-index-php/) > Probabilistic Occupancy Map <!-- [DeepSORT](https://github.com/nwojke/deep_sort) <br/> <br/> <br/> -->