Top-down, Spatio-Temporal Attentional Guidance for On-road Object Detection

On-road object detection is a crucial component for environmental perception in intelligent vehicles. Anchor generation is an intermediate step in object detection to derive a set of reference boxes for possible objects in different scales and aspect ratios. State-of-the-art object detectors rely on a massive number of pre-determined anchors over the whole scene. The required operational cost is a drawback in resource-constrained, mobile environments. In contrast, humans rapidly attend to relevant regions in the scene in detail based on the prior knowledge on the current goal and task in hand. Inspired by this observation, we aim to computationally model this top-down visual attention mechanism for the driving task to guide the anchoring process of on-road object detection. In particular, we use the knowledge about the environmental risk level and the underlying risk factors specifically for the driving task to derive an attention region to remain vigilant upon. Then, we perform anchor generation and subsequent operations for object detection only in the extracted attention region. Experimental results demonstrate that the proposed method significantly reduces the operational cost while preserving a competitive accuracy.