Object tracking in video sequences using information fusion principles. Meanshift kernel implementation using fuzzy rules

Abstract

Tracking of an object of interest is posed as a time varying problem and complexity of the tracking algorithm at any specific moment in time is dynamically adjusted based on the scene conditions as shown in fig 1. We have implemented kernel weighted discrete probability density function of the target color feature by using a Fuzzy Rule Based System. This technique based on Fuzzy Logic and Look-up tables not only facilitates near real time tracking in video feed but performs very competitively to the traditional mean shift algorithm which in its original form is very cumbersome to implement especially when a real time intervention is required for e.g. in surveillance or in safety critical applications where driving assistance is the prime objective. A centralized controller or supervisor utilizes Bhattacharyya quality control and standard conflict resolution techniques to decide the level of algorithmic complexity needed to resolve any potential differences of opinion that may arise among modules. By selecting appropriate loop as in the Fig 1 and hence the optimum level of information fusion our tracker becomes much more robust even during time spans when a large level of maneuverability is observed as compared to the traditional mean shift with an added advantage of real time capa.