Overview

  • Motivation
    • Real-time multi-person pedestrian tracking is a major component of many applications, e.g. autonomous driving or autonomous construction scenarios where pedestrians have to be detected and tracked in real time for safe operations.
    • Recently, a lot of works have tackled the problem of real-time single object tracking. Although achieving state-of-the-art accuracy while running as fast as hundreds of frames per second, these algorithms focus on tracking single object only, which is sub-optimal in realistic settings.
    • There are other recent methods that focus on multiple object tracking, which utilize a detection+tracking+assfication architecture. These methods achieve great performance but lack the ability to run in real time, which is essential to our work.
  • Problem
    • The main challenge is to handle some difficult scenarios, such as occlusion, appearance change, pedestrian crossing.
    • The second challenge is that the time complexity will increase with more tracking objects.
  • Solution
    • We are trying to use one-stage network for both detection and matching.
    • We believe Graph Neural Network (GNN) is useful for simultaneous detection and association
    • We designed a Non-Maximum Suppression specifically tailored for the tracking task