Evolving Head Tracking Routines With Brain Programming

Abstract

Visual tracking has long been studied in computer vision, and it has many practical applications such as surveillance and security, traffic control, human-computer interaction, and activity or behavior recognition to mention but a few. Head tracking attempts to follow a person's head within a video sequence. This paper presents a methodology that automatically designs an artificial dorsal stream for the problem of head tracking. Multiple visual operators are synthesized to obtain several visual and conspicuity maps that are fused into a saliency map, which is converted to a binary image, thus defining the proto-object. The automatic design of visual attention programs for the problem of head tracking is achieved through an optimization process following the Darwinian paradigm of artificial evolution. Artificial brains are synthesized using multiple visual operators embedded within a complex hierarchical procedure consisting of several key processes such as center-surround mechanisms, normalization, and pyramid-scale processes. The proposed strategy robustly handles many difficulties including occlusion, distraction, and illumination, and the resulting programs are real-time systems that are able to track a person's head with enough accuracy to automatically control the camera. Extensive experimentation shows that the proposed method outperforms several state-of-the-art methods in the challenging problem of head tracking.

Keywords

• Automatic programming
• brain modeling
• deep genetic programming
• deep learning
• learning
• optimization