Skip to content

Trackers

motcpp provides 9 state-of-the-art multi-object tracking algorithms. This guide helps you choose the right one for your application.

Overview

Tracker Type Speed Accuracy Use Case
SORT Motion ⚡⚡⚡⚡⚡ ⭐⭐⭐ Real-time, simple scenes
ByteTrack Motion ⚡⚡⚡⚡⚡ ⭐⭐⭐⭐ General purpose
OC-SORT Motion ⚡⚡⚡⚡ ⭐⭐⭐⭐ Occlusion handling
UCMCTrack Motion ⚡⚡⚡⚡ ⭐⭐⭐⭐ Camera motion
DeepOC-SORT ReID ⚡⚡⚡ ⭐⭐⭐⭐⭐ Re-identification
StrongSORT ReID ⚡⚡ ⭐⭐⭐⭐⭐ High accuracy
BoT-SORT ReID ⚡⚡ ⭐⭐⭐⭐⭐ Camera motion + ReID
BoostTrack ReID ⚡⚡ ⭐⭐⭐⭐⭐ State-of-the-art
HybridSORT ReID ⚡⚡ ⭐⭐⭐⭐⭐ Hybrid approach

Motion-Only Trackers

These trackers use only motion cues (Kalman filter + IoU matching) without appearance features.

SORT

Simple Online and Realtime Tracking — The foundational tracker that many others build upon.

#include <motcpp/trackers/sort.hpp>

motcpp::trackers::Sort tracker(
    0.3f,   // det_thresh: detection confidence threshold
    1,      // max_age: frames before track deletion
    50,     // max_obs: max observation history
    3,      // min_hits: hits before track confirmation
    0.3f    // iou_threshold: matching threshold
);

Best for: - Maximum speed requirements - Simple scenes with minimal occlusion - Baseline comparisons

Paper: arXiv:1602.00763 (2016)

ByteTrack

BYTE: Multi-Object Tracking by Associating Every Detection Box — Two-stage association for better handling of low-confidence detections.

#include <motcpp/trackers/bytetrack.hpp>

motcpp::trackers::ByteTrack tracker(
    0.3f,   // det_thresh
    30,     // max_age
    50,     // max_obs
    3,      // min_hits
    0.3f,   // iou_threshold
    false,  // per_class
    80,     // nr_classes
    "iou",  // asso_func
    false,  // is_obb
    0.1f,   // min_conf: low confidence threshold
    0.45f,  // track_thresh: high confidence threshold
    0.8f,   // match_thresh
    30,     // track_buffer
    30.0f   // frame_rate
);

Best for: - General-purpose tracking - Crowded scenes - When you need good speed/accuracy balance

Paper: ECCV 2022

OC-SORT

Observation-Centric SORT — Improved Kalman filter with observation-centric momentum.

#include <motcpp/trackers/ocsort.hpp>

motcpp::trackers::OCSort tracker(
    0.2f,    // det_thresh
    30,      // max_age
    50,      // max_obs
    3,       // min_hits
    0.3f,    // iou_threshold
    false,   // per_class
    80,      // nr_classes
    "iou",   // asso_func
    false,   // is_obb
    0.1f,    // min_conf
    3,       // delta_t: observation window
    0.2f,    // inertia: velocity dampening
    false,   // use_byte: use ByteTrack association
    0.01f,   // Q_xy_scaling: process noise
    0.0001f  // Q_s_scaling: scale noise
);

Best for: - Non-linear motion - Occlusion recovery - When objects temporarily disappear

Paper: CVPR 2023

UCMCTrack

Unified Camera Motion Compensation — Ground-plane tracking with camera parameters.

#include <motcpp/trackers/ucmc.hpp>

motcpp::trackers::UCMCTrack tracker(
    0.3f,    // det_thresh
    30,      // max_age
    50,      // max_obs
    3,       // min_hits
    0.3f,    // iou_threshold
    false,   // per_class
    80,      // nr_classes
    "iou",   // asso_func
    false,   // is_obb
    100.0,   // a1: confirmed association threshold
    100.0,   // a2: tentative association threshold
    5.0,     // wx: process noise x
    5.0,     // wy: process noise y
    10.0,    // vmax: max velocity
    0.033,   // dt: time step (1/fps)
    0.5f     // high_score: confidence split
);

Best for: - Drone/aerial footage - Moving cameras - When camera parameters are available

Paper: AAAI 2024

ReID-Enhanced Trackers

These trackers use appearance features (ReID embeddings) in addition to motion cues.

DeepOC-SORT

Deep OC-SORT — OC-SORT with deep appearance features.

#include <motcpp/trackers/deepocsort.hpp>

motcpp::trackers::DeepOCSort tracker(
    "osnet_x1_0.onnx",  // reid_weights
    false,              // use_half
    false,              // use_gpu
    0.2f,               // det_thresh
    30,                 // max_age
    // ... additional parameters
);

Best for: - Long-term tracking - Re-identification after long occlusions - Similar-looking objects

Paper: arXiv:2302.11813 (2023)

StrongSORT

StrongSORT++ — Enhanced SORT with appearance features and NSA Kalman.

#include <motcpp/trackers/strongsort.hpp>

motcpp::trackers::StrongSORT tracker(
    "osnet_x1_0.onnx",  // reid_weights
    false,              // use_half
    false,              // use_gpu
    0.3f,               // det_thresh
    30,                 // max_age
    // ... additional parameters
);

Best for: - High accuracy requirements - When appearance is distinctive - Benchmark evaluations

Paper: TMM 2023

BoT-SORT

BoT-SORT — Camera motion compensation with ReID.

#include <motcpp/trackers/botsort.hpp>

motcpp::trackers::BotSort tracker(
    "osnet_x1_0.onnx",  // reid_weights
    false,              // use_half
    false,              // use_gpu
    0.3f,               // det_thresh
    30,                 // max_age
    // ... additional parameters
);

Best for: - Moving cameras + appearance - Sports broadcasting - Surveillance with PTZ cameras

Paper: arXiv:2206.14651 (2022)

BoostTrack

BoostTrack — Latest SOTA with confidence-aware tracking.

#include <motcpp/trackers/boosttrack.hpp>

motcpp::trackers::BoostTrackTracker tracker(
    "osnet_x1_0.onnx",  // reid_weights
    false,              // use_half
    false,              // use_gpu
    0.3f,               // det_thresh
    30,                 // max_age
    // ... additional parameters
);

Best for: - Maximum accuracy - Research/benchmarks - When speed is not critical

Paper: MVA 2024

HybridSORT

HybridSORT: A Simple but Strong Association for Multi-Object Tracking — Combines motion and appearance cues with hybrid matching.

#include <motcpp/trackers/hybridsort.hpp>

motcpp::trackers::HybridSort tracker(
    0.3f,   // det_thresh
    30,     // max_age
    50,     // max_obs
    3,      // min_hits
    0.3f,   // iou_threshold
    "hmiou" // asso_func: "iou" or "hmiou"
);

Best for: - Balanced accuracy and speed - Scenes with moderate occlusion - When ReID is available

Paper: arXiv:2303.XXXX

Choosing a Tracker

                    ┌─────────────────────┐
                    │ Need maximum speed? │
                    └──────────┬──────────┘
                               │
              ┌────────────────┴────────────────┐
              │                                 │
              ▼                                 ▼
         ┌────────┐                       ┌────────┐
         │  Yes   │                       │   No   │
         └───┬────┘                       └───┬────┘
             │                                │
             ▼                                ▼
    ┌─────────────────┐          ┌─────────────────────┐
    │ SORT/ByteTrack  │          │ Need ReID features? │
    └─────────────────┘          └──────────┬──────────┘
                                            │
                          ┌─────────────────┴─────────────────┐
                          │                                   │
                          ▼                                   ▼
                     ┌────────┐                          ┌────────┐
                     │   No   │                          │  Yes   │
                     └───┬────┘                          └───┬────┘
                         │                                   │
                         ▼                                   ▼
              ┌──────────────────┐              ┌──────────────────┐
              │ OC-SORT/UCMCTrack│              │ StrongSORT/      │
              │ (motion focus)   │              │ BoostTrack       │
              └──────────────────┘              └──────────────────┘

Benchmarks

Performance on MOT17 test set:

Tracker HOTA↑ MOTA↑ IDF1↑ FPS
SORT 62.4 75.2 69.2 1250
ByteTrack 66.5 76.4 77.6 1100
OC-SORT 64.6 73.9 74.4 850
UCMCTrack 64.0 75.6 73.9 980
DeepOC-SORT 65.8 75.1 76.2 120
StrongSORT 66.2 75.8 77.1 95
BoT-SORT 66.8 76.2 78.3 85
BoostTrack 67.5 77.1 79.2 75