PROCEEDINGS ARTICLE | February 20, 2018
Proc. SPIE. 10697, Fourth Seminar on Novel Optoelectronic Detection Technology and Application
KEYWORDS: Visual process modeling, Video acceleration, Detection and tracking algorithms, Sensors, Video, Feature extraction, Computer vision technology, Machine vision, Optical tracking, Neural networks, Target recognition
The paper proposes a visual tracking method based on deep learning without online model updating. In consideration of the advantages of deep learning in feature representation, deep model SSD (Single Shot Multibox Detector) is used as the object extractor in the tracking model. Simultaneously, the color histogram feature and HOG (Histogram of Oriented Gradient) feature are combined to select the tracking object. In the process of tracking, multi-scale object searching map is built to improve the detection performance of deep detection model and the tracking efficiency. In the experiment of eight respective tracking video sequences in the baseline dataset, compared with six state-of-the-art methods, the method in the paper has better robustness in the tracking challenging factors, such as deformation, scale variation, rotation variation, illumination variation, and background clutters, moreover, its general performance is better than other six tracking methods.
