Integration of Mean-Shift and Particle Filter: A Survey

Abstract

Object tracking has become the cornerstone of many computer vision applications. Numerous object tracking methods have surfaced in the research community which are intended for high level applications such as automatic data analysis for activity recognition. Most of the methods are either too constrained in the context of the given application or they are costly in terms of computations to meet the real-time requirements. For example, Mean-Shift (MS) has rose to prominence due to its ease of implementation and robustness to various deformations however it fails to track objects with small sizes, fast motion and full occlusion. On the other hand, Particle Filter (PF) is known for its efficiency, accuracy and robustness to small sizes, fast object motion and full occlusion however it is heavily influenced by the number of particles, besides the sample degeneracy and impoverishment problems. Decoupling the disadvantages of both the methods gave birth to a new era of modern trackers known as hybrid systems that are more efficient, accurate and robust to the aforementioned constraints.A few survey papers on object tracking has been published in the scientific circles during the last decade however we feel that this popular integration of MS and PF is still unregistered.

Adaptive Framework for Multi-Feature Hybrid Object Tracking

“…≤ N T we redistribute the clusters of particles, so that new particles are generated out of this cluster and re-initialize their weights to 1 N s . N T is a threshold that triggers re-sampling if it is greater than N e…”

“…The increase in the computational power of existing systems has led to huge investments in automated data analysis. Object tracking is one such class of algorithms that automatically locates the region of interest, possibly obscured by challenging constraints [1]. These constraints are what defines the requirements that should be considered while developing real-time robust object tracking algorithms.…”

A Hybrid Tracking System of Full-Body Motion Inside Crowds

“…Thus, the absolute positions for our studies on pedestrian dynamics and for the proposed method are provided by a well-calibrated overhead camera grid [ 7 ]. For the detection and tracking of marked heads within image sequences, a lot of well established methods with low error exist (e.g., mean-shift and particle filter [ 53 , 54 ]). For the structured marker we use for detection the shape of oriented isolines of the same brightness, and for tracking the iterative Lucas Kanade feature tracker [ 55 ].…”