Low-cost multiple object tracking for embedded vision applications

Abstract

This paper presents a low-cost multiple object tracking (MOT) technique by employing a novel appearance update model for object appearance modeling using K-means. The state-of-the-art work has attained a very high accuracy without considering the real-time aspects necessitated by currently trending embedded vision platforms. The major research on multiple object tracking is used to update the appearance model in every frame while discounting its persistent nature. The proposed appearance update model reduces the computational cost of the state-of-the-art MOT 6-fold by exploiting this facet of persistent appearance over the sequence of frames. To ensure accuracy, the proposed model is tested on different publicly available standard datasets with challenging situations for both indoor and outdoor scenarios. The experimental results illustrate that our model successfully achieves multiple object tracking while coping with long-term and complete occlusion. The proposed method achieves the same accuracy in comparison with the state-of-the-art baseline methods. Moreover, and most importantly, the proposed method is cost-effective in terms of computing and/or memory requirements in comparison to the state-of-the-art techniques. All these traits make our design very suitable for real-time and embedded video surveillance applications with low computing/memory resources.

Real-time motion tracking enhancement via data-fusion based particle filter

“…At the beginning, a smaller window containing Kinect's output position for object's region of interest is identified in order to form a basis for performing further operations within a quite smaller solution space rather than whole stage image window Kinect-acquired RGB images are converted to 8-bit grayscale and Depth images are quantized into 8-bit images in order to decrease computational load. Active window images are then subtracted from the corresponding background images to obtain moving only regions as indicated in the network of Shehzad et al [27]. In the next step, a separate cue which means an extraction of requested object coordinates is achieved by using the RGB and depth moving only active window images while Kinect's output is directly taken as a distinct cue in this phase.…”