FORT: Fisheye Online Realtime Tracking with an Improved Kalman Filter

The goal of human tracking is to detect people in a scene and assign them a unique identifier that the tracker will follow across multiple frames. Our tracker, FORT, implements deep learning solutions such as, YOLOv7 for detection, ResNeXt-50 for feature extraction and re-identification and an adapted Kalman filter for tracking. The goal is to present a real time tracking solution for the complex environment of top-view, fisheye images. The proposed solution is then compared with the BoT-SORT and StrongSORT trackers on a custom fisheye Multiple Object Tracking (MOT) Challenge dataset.

I. Introduction

Human tracking is a still current challenging task [1]. In a video, people change direction, pass next to each other, are occluded, their detected positions can be unreliable and as such, it is easy to lose or not be able to find a person. These problems are emphasised in distorted contexts such as in top view fisheye cameras. Indeed, a fisheye camera increases the focal distance and embraces a wide field of view [2], [3], making people look inclined and distorted in the image. Consequently, people have irregular movements in the image, their aspect ratio, size and orientation can change drastically as they move. This context is extensively studied for detection [4]–[9] and tracking tasks [10]–[13] as fisheye cameras give 180° of vision and as such find applications in many fields such as security, autonomous driving and aeronautics.