Evaluating image features in real-world scenarios

The rapid growth in the autonomous vehicle technology field has triggered multiple advances in feature-based computer vision applications. However, most applications are based on using optimal data, for example traffic data under sunny conditions and regularly do not include adverse weather or low-light conditions. This paper presents a survey describing the effect of real-world perturbations on the performance of 6 conventional feature-based methods and 2 convolutional neural network (CNN) methods in 3 scenarios: during one year, during one month and all-year weather and illumination changes. An evaluation is made on the suitability of the selected methods, using the amount of extracted inliers and processing times as metrics for a low-, medium and high-complexity case. It is shown that perturbations real-world conditions exceed the performance differences between conventional methods and that a performance increase is achieved with a well-tuned CNN at the expense of one order of magnitude of processing time.