Another difference from naturalistic driving is that drivers don`t always look in the front direction (i.e. looking in the direction of F1 view) as was the case in the videos used. With regard to the longer duration, it can be expected that changing this function to a more realistic display behavior will reduce the accuracy of the annotations due to the greater confusion between the regions considered. Manufacturer accuracy = [number of correct errors / total number of correct errors and omissions] > I have the data set and test and train samples (in a ratio of 30:70). > Now I`ve created a model with a logistic regression, i.e. LRM1, and a calculated accuracy, which looks good. > predicted on the test set with the LRM1 model > The ROC curve was displayed on the train dataset and the new cut-off point was determined. Based on the new cut-off point, the classification was performed on the model predicted by the test and the accuracy was calculated. Gamer, M., Lemon, J. & Singh, I. F. P. (2019).
IRR: Different coefficients of reliability and agreement of the evaluator. CRAN.R-project.org/package=irr. If larger clusters can be used, greater accuracy can be achieved. For example, if all the locations of the cabin view are grouped into two categories (top and bottom) and the rest of the regions are retained, an overall accuracy of 75% is achieved. This accuracy may still be lower than what is required for a naturalistic driving study or the use of annotated data to develop automatic detection algorithms. One possible reason may be confusion with observation positions that are not part of the cabin area, such as .B. between the upper centre of the cab (C7) and the cab at the top right (C8) on the one hand and the front rear-view mirror (F4) and the right pedestrian rear-view mirror (R4) on the other hand. Overall, it must be concluded that the accuracy of the annotation of the off-road gaze of (1) the viewing positions selected in the study depends on (2) the grouping of the viewing positions and (3) possibly the size of the driver. For street views, the accuracy for the F2 (center of the front window) and F3 (right front window) positions was less than 25%, as these positions were often confused with straight side mirrors. Even if the F2 and F3 visualization positions are grouped together, the resulting accuracy is still less than 50%.
The confusion of these places was unidirectional: the right mirrors were never confused with F2 and only occasionally with F3. This indicates an exceedance effect, as described by Moors et al. (2015). However, the data also shows that the viewing directions to the left and right can be distinguished with great precision (the confusion between left and right accounts for only 1.12% of all responses in the left and right areas). For studies that only distinguish between the directions of the left and right gaze (e.g.B. infant studies from the gaze), the current results should therefore not be of concern. Naturalistic driving studies often use manual annotations to determine the driver`s current viewing direction based on video images. When using such an annotation, it is often assumed that the annotators are correct if the annotators agree. In this study, we tested this hypothesis by showing videos of truck drivers looking at predefined regions inside and outside the truck and asking participants to name those regions. Our five main results are as follows: 1) there are large differences in accuracy and percentage of agreement between viewing positions, 2) some observation positions show a significant influence of the cyclist`s height on accuracy, 3) average accuracy was consistently low on annotators, 4) some observation positions showed very low accuracy despite a moderate degree of agreement, and 5) Grouping viewing positions into larger viewing areas improves accuracy. Next, we discuss the implications of our findings, followed by the limitations of the study and recommendations for future research.
Another limitation may be that, unlike naturalistic conduct, the fixation interval of 2 s in our study may have been quite long (knowing that the human gaze occurs about three times per second). For annotation, a relatively long attachment time can be expected to improve the accuracy of the annotation. Therefore, given that the accuracy of the annotation was rather low, we expect that the accuracy of the annotation for actual naturalistic driving data could be even lower. The ability to determine another person`s line of sight was an area of general interest. Studies on this topic suggest that the accuracy of such a perception depends on where the observed person is looking. .