An oriented gradient based image quality metric for pedestrian detection performance evaluation

Full reference image quality metrics are important tools for optimizing system design parameters associated with image acquisition, compression and transmission. While optimizing systems for perceptual quality is important, in the automotive environment Advanced Driver Assistance Systems (ADAS) such as automated pedestrian detection are becoming a common feature of in-vehicular vision systems. As such, automotive image quality must also be tuned for optimal machine vision performance. In this paper the effects of transmission artifacts on the performance of a number of state-of-the-art pedestrian detection algorithms are evaluated. We demonstrate that the human visual system may not perceive distortions that adversely affect machine vision performance. As a result, existing full-reference image quality metrics are not necessarily accurate predictors of machine vision performance on transmitted video sequences. To address this problem, a novel, computationally inexpensive, full-reference objective quality metric based on histogram of oriented gradients is proposed. The proposed metric accurately predicts algorithm performance in the presence of transmission artifacts. The metric can be used at the system design stage in order to optimize image capture parameters for machine vision performance without the need for annotated test databases, which are both expensive and time consuming to produce.