Most of the sequential importance resampling tracking algorithms use arbitrarily high number of particles to achieve better performance, with consequently huge computational costs. This article aims to address the problem of occlusion which arises in visual tracking, using fewer number of particles. To this extent, the mean-shift algorithm is incorporated in the probabilistic filtering framework which allows the smaller particle set to maintain multiple modes of the state probability density function. Occlusion is detected based on correlation coefficient between the reference target and the candidate at filtered location. If occlusion is detected, the transition model for particles is switched to a random walk model which enables gradual outward spread of particles in a larger area. This enhances the probability of recapturing the target post-occlusion, even when it has changed its normal course of motion while being occluded. The likelihood model of the target is built using the combination of both color distribution model and edge orientation histogram features, which represent the target appearance and the target structure, respectively. The algorithm is evaluated on three benchmark computer vision datasets: OTB100, V OT18 and TrackingNet. The performance is compared with fourteen state-of-the-art tracking algorithms. From the quantitative and qualitative results, it is observed that the proposed scheme works in real-time and also performs significantly better than state-of-the-arts for sequences involving challenges of occlusion and fast motions.
- object tracking
- particle filter