Object Cueing Using Biomimetic Approaches to Visual Information Processing

Thousands of years of evolution have produced the human vision system that computers cannot replicate well. Humans are still unsurpassed in their ability to search for objects in visual scenes. To successfully detect objects in cluttered scenes, the human brain is thought to rely on multiple factors: prior probabilities of object occurrence, global scene statistics and object co-occurrence. Machine object detection continues to be one of the hot areas of computer vision, and it attests both to the fundamental importance of the problem and the fact that state-of-the-art algorithms are shy of performing well in practice. We find that the key problems in current FMV data recognition frameworks (speed, accuracy, robustness, clutter, occlusion, change in view) can be overcome by maturing computer vision systems with architectures and components inspired by advances in neuroscience and computational human vision. In Phase I, we have demonstrated a clear path towards the solving small object recognition problem in overhead videos. In Phase II, we will focus on delivering end-to-end biomimetic solution to scene and object recognition with high sensitivity and specificity for a range of overhead video feeds from moving aerial platforms, from fast-moving narrow field sensor feeds to persistent wide area feeds.