Robust Detection and Tracking of Long-range Target in a Compound Framework

Abstract

In this paper, we propose a compound framework that combines detection and tracking for long-range target tracking. Heavily modified Ferns with FAST-9 is firstly designed for target detection and initialization for tracker. During tracking phase, detector hands over tracking task to a new learning-based tracker called Tracking with Multi-resolution Prior (TMRP) that we present. The main advantage of our TMRP algorithm is the employment of offline training stage to remove most computational cost from online tracking phase, which means much higher real-time performance. During training phase of TMRP, we divide the large motion range into multiple levels artificially according to the law of “coarse-to-fine”, then employ basic training equations to obtain multi-resolution priori error Jacobian matrices. This step can remarkably improve the performance of tracking violent movements. For our special application of long-range tracking, we design a multi-template-relay tactics to handle great variations of target scale while moving closer to target from faraway. We describe the real-time implementation our framework and TMRP. At last, an extensive quantitative experiments demonstrate that our method works well both in robustness and real-time performance.