Multi-Object tracking based on Kalman Filtering Combining Radar and Image Measurements

The purpose of this paper is to develop a tracking system. The designed platform is based on two kinds of physical sensors: a doppler radar module and HD Camera. All the measurements from those modules are processed using a data fusion method. In this work, a method based on the Gaussian mixture model is used for foreground detection (i.e. background subtraction). After that our move to a filtering step which is used to refine the detection results firstly obtained, then a tracking process is introduced. As a final stage, all the vision-based measurements are combined with the processed radar raw data. Here, the goal is to perfectly estimate the target velocity in the real time. Added to target 2D positions, this speed information is considered as a third dimension. This is very useful in many applications such as traffic control, robotics, autonomous vehicles etc. In this work a set of experiments is conducted in order to validate the developed tracking method.

I. Introduction

Multi Object detection and tracking are among the most studied issues in recent years. Nowadays, multi-object detection and tracking in a dynamic environment, especially for humans and vehicles, is one of the current challenging research topics in computer vision. It is a key technology to fight against terrorism, crime, efficient management of traffic control and many industrial applications. It reduces the quantity of information to be analyzed. This task aims to separate the objects constituting the frames of video sequence into moving and fixed objects. In this step, there are several approaches for the detection of moving objects such as the Background-based approach and Frame-based approach. Background-based approach involves studying each pixel of the image and compare it with a reference image or a model of our background. If the difference is greater than a fixed threshold, the pixel is classified as foreground, otherwise it is classified as background [1] [2] [3]. Frame-based approach are based on the difference between consecutive pictures and they dont require a background models, the extraction of the foreground is done by the analysis of the temporal variation of the pixel intensity [4] [5]. After the detection step, a tracking step is necessary. Tracking is defined as a spatiotemporal location of moving objects in a video. It can be performed with several techniques. In literature, there are several offered examples to show the principle of operation, the advantages and the limits of each method [6]. Referring to many research works, depending on the type of application, there are three classes of object tracking approaches: •

define the target positions in 2D image plan (with a single camera) [7] [8],

•

define the target positions in 3D image plan (stereo vision, range data fusion) [9] [10],

•

find out positions (2D or 3D) and velocity (single camera, speed data fusion) [11] [12].