The document compares four tracking algorithms (SORT, DeepSORT, ByteTrack, OC-SORT) for multi-object tracking, explains tracking-by-detection, and outlines a workflow using RF-DETR in Roboflow to test and deploy trackers. It covers how each algorithm predicts motion, associates detections, and manages object tracks, ending with a detailed SORT procedure using Kalman filtering and the Hungarian algorithm.
Build a real-time video analytics pipeline using YOLO26 for detection and ByteTrack for tracking, with line-crossing counting and a live visualization. The system treats detection, tracking, and analytics as separate stages, enabling persistent identities and time-based events.
Sky Tracker is a CPU-based object tracker for drones and fast aerial objects. It offers box-based targeting, frame-by-frame tracking with optional small neural net, cross-platform support, multi-language packaging, a free demo, and a one-time license model for Personal and Commercial use.
This document explains how object re-identification (Re-ID) links the same person or vehicle across cameras and time gaps using appearance embeddings, detailing the detection, encoding, and matching steps, training losses, and its relation to tracking and facial recognition, plus practical applications in retail and traffic analysis.
A Fujitsu researcher summarizes the IEEE VTC2026-Fall conference and presents a real-time cross-modal fusion approach that combines mmWave radar and low-resolution infrared sensing to improve robust multi-object tracking, addressing radar dropouts and thermal noise while staying privacy-preserving.
SAM 3 runs offline on a parking-lot video, segmenting all 121 cars per frame from a single 'car' prompt, while locally mapping 221 spots and tracking vehicles frame-by-frame; the system updates a free-spot counter as cars move, with no license plates read.
ReLift treats dense 3D tracking as a trajectory repair problem, using a geometry-only, feed-forward framework to fix and complete lifted trajectories. It achieves strong accuracy on WorldTrack, outperforming baselines and offering speedups without retraining.
ReLift reframes dense 3D tracking as geometric trajectory repair, using a trajectory corrector to fix and complete lifted trajectories with only geometric inputs, achieving state-of-the-art results on WorldTrack without retraining and showing strong generalization across trackers and geometry models.
tamserve is a Python package that runs SAM-family video trackers across multiple cameras using one model copy per GPU. Its reported EdgeTAM tests show lower memory use and higher total throughput than separate model copies, with nearly unchanged mask accuracy.
The document compares image and video processing modes. Image calls treat each still image independently, while video mode tracks every decoded frame in order.