Attention-Based Vision for Autonomous Vehicles

Humans are continually subjected to an overwhelming stream of visual, auditory, olfactory, and touch stimuli. We operate effectively because we are capable of focusing attention on the relevant items, and capable of accurately perceiving tiny discrepancies that are most relevant to our survival while ignoring broad swaths of superfluous data. Our ability to quickly key in on the most relevant sensory data can be shown clearly in, for example, measurements of eye-saccade patterns: when presented with an image of a face, humans focus attention on the most informative parts of an expression, the eyes and mouth, which provide important cues as to the intention and attitude of the individual. Current autonomous systems lack this ability to rapidly focus on the most critical data, and suffer several deficiencies as a result. First, they are too brittle: they fail to adapt effectively to changing requirements and conditions typical of military applications. Second, they are too slow: Third, they make mistakes, failing to handle simple navigation and recognition tasks easily handled by humans. Drawing inspiration from recent understanding in human cognition and sensor processing, we propose SACCADE, a system to rapidly identify and process sensor data from the most relevant objects