Abstract:Multi-target tracking technology can achieve real-time tracking by detecting the identity information of multiple targets. It is a key technology in the field of computer vision and has played an excellent role in the fields of artificial intelligence such as military, security, and autonomous driving.However, current multi-target tracking algorithms cannot accurately identify and track targets in scenes where targets are densely distributed or there are many occlusions, and the tracking results obtained contain a large number of false negatives and false positives. In order to solve the above problems, we propose a multi-target tracking algorithm based on dual-branch feature enhancement and multi-level trajectory correlation.This algorithm uses a dual-branch feature learning network to learn the particularity and correlation of detection and tracking tasks, which alleviates excessive competition between the two tasks and can extract sufficient target feature information.This algorithm introduces an association matrix, uses more timing information to predict offset vectors, and reduces the number of identity switching times. We adopt a multi-level trajectory association strategy and re-divide detection frames into high-scoring frames and low-scoring frames, and use different matching methods to associate with trajectories to reduce the number of missing trajectories.On the typical multi-target tracking data sets MOT17 and MOT20, we conducted comparative experiments on 6 related algorithms such as JDE and CenterTrack. Experimental results show that the MOTA and IDF1 values of the proposed algorithm on MOT17 reached 68.2% and 68.5% respectively, which were improved by 2.1% and 4.3% respectively compared with the benchmark algorithm CenterTrack.On MOT20, the MOTA and IDF1 values of our algorithm reached 52.7% and 48.2%, which were 1.4% and 7.9% higher than the baseline algorithm respectively. All these results show that the proposed algorithm achieves excellent tracking performance in complex scenes.

Three-dimensional Multi-Target Tracking of Point Cloud from Bird's-Eye View

Online Multi-Object Tracking from A Bird's-Eye View by Fusion of Millimeter-Wave Radar and Vision

Point Cloud Multi-view 3D Target Detection Method Based on Complementarity

Object-Level Pseudo-3D Lifting for Distance-Aware Tracking

3D Multi-Object Tracking in Point Clouds Based on Prediction Confidence-Guided Data Association

EasyTrack: Efficient and Compact One-stream 3D Point Clouds Tracker

Time-Sensitive 3D Single Target Tracking Based on Deep Hough Optimized Voting

3D Single Object Tracking Network Based on Point Cloud Pre-segmentation

Tracking Objects as Points

PointTrackNet: An End-to-End Network For 3-D Object Detection and Tracking From Point Clouds

Factor Graph based 3D Multi-Object Tracking in Point Clouds

DeepPCT: Single Object Tracking in Dynamic Point Cloud Sequences

A Dynamic 3D Multi-Object Tracking Method Based on Spatiotemporal Features

3D Multi-Object Online Tracking with Multi-View Clustering

A Multi-target Tracking Algorithm Based on Multiple Cameras

Beyond 3D Siamese Tracking: A Motion-Centric Paradigm for 3D Single Object Tracking in Point Clouds

A Multi-Object Tracking Algorithm Based on Multi-Camera

A Tracking-By-Detection Based 3D Multiple Object Tracking for Autonomous Driving

A Lightweight and Detector-Free 3D Single Object Tracker on Point Clouds

Multi-target tracking based on appearance features and similarity fusion

Real-time Multi-Object Tracking Based on Bi-directional Matching