Image Based Multi Objects' Trajectories Estimation through Reduced Order Observer

Abstract

Simple and robust estimation algorithm for real-time multiple object detection and tracking without relying on classification through appearance is presented. Detection of the object in low contrast thermal Infrared (IR) video sequence is performed through adaptive thresholding and image segmentation. Probabilistic data association of detected objects linked with estimation technique is then modeled into a network having the constraint of non-overlapping trajectories. The proposed framework is based on object tracking through MAP linked with reduced order observer (ROO) for robust estimation of trajectories that uses the spatial location, appearance and velocity of objects as states. Future estimation efficiency increased by incorporating prediction error. Computational complexity is reduced by reusing the calculations, instead of applying a Single object tracking model multiple times to achieve the task of multiple object tracking. Identity inference and occlusion are handled intrinsically in the algorithm.