Explicit Image Quality Detection Rules for Functional Safety in Computer Vision

Increasing Safety by Combining Multiple Declarative Rules in Robotic Perception Systems

“…Initial steps towards establishing a safe implementation of perception systems have been demonstrated using explicit declarative rules to address specific issues in a perception pipeline [13,14]. The rules are focused on different particularities of an image such as pixel distributions, pixel changes, and frequency.…”

“…The many categories however also makes the classification problem is harder and thereby the performance of the algorithm deteriorates. As a result the concept of soft-boundaries was introduced to evaluate multiclass classification systems without penalizing the system excessively [14]. The computer vision rules have been implemented in the Vision Safety Language (ViSaL) DSL for automatic generation of their implementation [13] and an initial assessment has been performed with respect to readability with the goal of facilitating certification [12].…”

“…In our experiments we use datasets consisting of images labelled using a usability category, where 5 is a usable image and 1 is unusable, i.e., unusable to make reliable safety decisions. Datasets consists of rgb and depth information, and are processed by the rules ("raw data"), based on which thresholds are found in the data using precisionrecall curves ("threshold data") [14].…”

“…A requirement for introducing autonomy in established domains is safety, which is done through compliance with functional safety standards that rely on code and documentation reviews. To support the use of vision in safety-critical systems, we propose to use simple and explicit computer vision rules as a means to determine specific problems in input images and the perception system as a whole [14]. Increasing the readability of the code and the perception rules can increase the overall trustworthiness of the system [12], facilitating the certification process [14].…”

“…To support the use of vision in safety-critical systems, we propose to use simple and explicit computer vision rules as a means to determine specific problems in input images and the perception system as a whole [14]. Increasing the readability of the code and the perception rules can increase the overall trustworthiness of the system [12], facilitating the certification process [14]. Such simple and explicit computer vision rules can be implemented in a domain specific programming languages (DSL), supporting safe implementation of vision for safety critical systems [13].…”