The detection and recognition of objects is essential for the operation of autonomous vehicles and robots. Designing and predicting the performance of camera systems intended to supply information to neural networks and vision algorithms is nontrivial. Optimization has to occur across many parameters, such as focal length, f-number, pixel and sensor size, exposure regime and transmission schemes. As such numerous metrics are being explored to assist with these design choices. Detectability index (SNRI) is derived from signal detection theory as applied to imaging systems and is used to estimate the ability of a system to statistically distinguish objects [1], most notably in the medical imaging and defense fields [2]. A new metric is proposed, Contrast Signal to Noise Ratio (CSNR), which is calculated simply as mean contrast divided by the standard deviation of the contrast. This is distinct from contrast to noise ratio which uses the noise of the image as the denominator [3,4]. It is shown mathematically that the metric is proportional to the idealized observer for a cobblestone target and a constant may be calculated to estimate SNRI from CSNR, accounting for target size. Results are further compared to Contrast Detection Probability (CDP), which is a relatively new objective image quality metric proposed within IEEE P2020 to rank the performance of camera systems intended for use in autonomous vehicles [5]. CSNR is shown to generate information in illumination and contrast conditions where CDP saturates and further can be modified to provide CDP-like results.
As autonomous vehicles and machines, such as self-driving cars, agricultural drones and industrial robots, become ubiquitous, there is an increasing need to understand the objective performance of cameras to support these functions. Images go beyond aesthetic and subjective roles as they assume increasing aspects of control, safety, and diagnostic capabilities. Radiometry and photometry are fundamental to describing the behavior of light and modeling the signal chain for imaging systems, and as such, are crucial for establishing objective behavior. As an engineer or scientist, having an intuitive feel for the magnitude of units and the physical behavior of components or systems in any field improves development capabilities and guards against rudimentary errors. Back-of-the-envelope estimations provide comparisons against which detailed calculations may be tested and will urge a developer to “try again” if the order of magnitude is off for example. They also provide a quick check for the feasibility of ideas, a “giggle” or “straight-face” test as it is sometimes known. This paper is a response to the observation of the authors that, amongst participants that are newly relying on the imaging field and existing image scientists alike, there is a general deficit of intuition around the units and order of magnitude of signals in typical cameras for autonomous vehicles and the conditions within which they operate. Further, there persists a number of misconceptions regarding general radiometric and photometric behavior. Confusion between the inverse square law as applied to illumination and consistency of image luminance versus distance is a common example. The authors detail radiometric and photometric model for an imaging system, using it to clarify vocabulary, units and behaviors. The model is then used to estimate the number of quanta expected in pixels for typical imaging systems for each of the patches of a MacBeth color checker under a wide variety of illumination conditions. These results form the basis to establish the fundamental limits of performance for passive camera systems based both solely on camera geometry and additionally considering typical quantum efficiencies available presently. Further a mental model is given which will quickly allow user to estimate numbers of photoelectrons in pixel.
Autonomous vehicles rely on the detection and recognition of objects within images to successfully navigate. Design of camera systems is non-trivial and involves trading system specifications across many parameters to optimize performance, such as f-number, focal length, CFA choice, pixel, and sensor size. As such, tools are needed to evaluate and predict the performance of such cameras for object detection. Contrast Detection Probability (CDP) is a relatively new objective image quality metric proposed to rank the performance of camera systems intended for use in autonomous vehicles. Detectability index is derived from signal detection theory as applied to imaging systems and is used to estimate the ability of a system to statistically distinguish objects, most notably in the medical imaging and defense fields. A brief overview of CDP and detectability index is given after which an imaging model is developed to compare and explore the behavior of each with respect to camera parameters. Behavior is compared to matched filter detection performance. It is shown that, while CDP can yield a first order ranking of camera systems under certain constraints, it fails to track detector performance for negative contrast targets and is relatively insensitive.