RESUMEN
In recent years, significant advances have been made in the development of Advanced Driver Assistance Systems (ADAS) and other technology for autonomous vehicles. Automated object detection is a crucial component of autonomous driving; however, there are still known issues that affect its performance. For automotive applications, object detection algorithms are required to perform at a high standard in all lighting conditions; however, a major problem for object detection is poor performance in low-light conditions due to objects being less visible. This study considers the impact of training data composition on object detection performance in low-light conditions. In particular, this study evaluates the effect of different combinations of images of outdoor scenes, from different times of day, on the performance of deep neural networks, and considers the different challenges encountered during the training of a neural network. Through experiments with a widely used public database, as well as a number of commonly used object detection architectures, we show that more robust performance can be obtained with an appropriate balance of classes and illumination levels in the training data. The results also highlight the potential of adding images obtained in dusk and dawn conditions for improving object detection performance in day and night.
RESUMEN
Image decolorization is an image pre-processing step which is widely used in image analysis, computer vision, and printing applications. The most commonly used methods give each color channel (e.g., the R component in RGB format, or the Y component of an image in CIE-XYZ format) a constant weight without considering image content. This approach is simple and fast, but it may cause significant information loss when images contain too many isoluminant colors. In this paper, we propose a new method which is not only efficient, but also can preserve a higher level of image contrast and detail than the traditional methods. It uses the information from the cumulative distribution function (CDF) of the information in each color channel to compute a weight for each pixel in each color channel. Then, these weights are used to combine the three color channels (red, green, and blue) to obtain the final grayscale value. The algorithm works in RGB color space directly without any color conversion. In order to evaluate the proposed algorithm objectively, two new metrics are also developed. Experimental results show that the proposed algorithm can run as efficiently as the traditional methods and obtain the best overall performance across four different metrics.
RESUMEN
In advanced driver assistance systems (ADAS) or autonomous vehicle research, acquiring semantic information about the surrounding environment generally relies heavily on camera-based object detection. Image signal processors (ISPs) in cameras are generally tuned for human perception. In most cases, ISP parameters are selected subjectively and the resulting image differs depending on the individual who tuned it. While the installation of cameras on cars started as a means of providing a view of the vehicle's environment to the driver, cameras are increasingly becoming part of safety-critical object detection systems for ADAS. Deep learning-based object detection has become prominent, but the effect of varying the ISP parameters has an unknown performance impact. In this study, we analyze the performance of 14 popular object detection models in the context of changes in the ISP parameters. We consider eight ISP blocks: demosaicing, gamma, denoising, edge enhancement, local tone mapping, saturation, contrast, and hue angle. We investigate two raw datasets, PASCALRAW and a custom raw dataset collected from an advanced driver assistance system (ADAS) perspective. We found that varying from a default ISP degrades the object detection performance and that the models differ in sensitivity to varying ISP parameters. Finally, we propose a novel methodology that increases object detection model robustness via ISP variation data augmentation.
RESUMEN
Interacting with other roads users is a challenge for an autonomous vehicle, particularly in urban areas. Existing vehicle systems behave in a reactive manner, warning the driver or applying the brakes when the pedestrian is already in front of the vehicle. The ability to anticipate a pedestrian's crossing intention ahead of time will result in safer roads and smoother vehicle maneuvers. The problem of crossing intent forecasting at intersections is formulated in this paper as a classification task. A model that predicts pedestrian crossing behaviour at different locations around an urban intersection is proposed. The model not only provides a classification label (e.g., crossing, not-crossing), but a quantitative confidence level (i.e., probability). The training and evaluation are carried out using naturalistic trajectories provided by a publicly available dataset recorded from a drone. Results show that the model is able to predict crossing intention within a 3-s time window.