A Dataset for Provident Vehicle Detection at Night

Provident vehicle detection at night for advanced driver assistance systems

“…For vehicle detection at night, it is difficult to judge which method is superior to another as the domain has not agreed on a benchmark dataset like the KITTI benchmark for daylight (Geiger et al, 2012), which is also criticized by other authors (Sun et al, 2006, Juric & Loncaric, 2014-authors reported results with around 90 % accuracy and small error rates for both rule-based and NN-based classifiers on their private datasets (e. g., compare the evaluations of Mo et al, 2019 andSatzoda &Trivedi, 2019). Nevertheless, it has to be expected that rule-based methods are superior if computational complexity constraints apply (e. g., see the number of parameters and the number of GFLOPs in Table 2 of Saralajew et al, 2021) and if somebody wants to use the detections for high-stakes decisions (Rudin, 2019). Therefore, as the scope of this work is to apply the detection algorithm in a test vehicle for a driver assistance system, we focus on a rule-based proposal generation algorithm with a shallow NN on top to classify the proposals (whereas the NN is not specific to our approach and can be replaced by any other classification method).…”

“…Whether this result is partly due to the simplicity of their bounding box generation, perhaps causing unexpected biases, is unclear. Saralajew et al, 2021 extended the work of Oldenziel et al, 2020 and published the PVDN dataset, the first containing approximately 60 K driver assistance camera images (grayscale) annotated by keypoints for the task to providently detect oncoming vehicles at nighttime. Together with the dataset published by Bell et al, 2021, these two datasets are the only large-scale datasets publicly available for the detection of vehicles at nighttime and annotated by keypoints-other available datasets for this task use bounding boxes (e. g., Rezaei & Klette, 2017, Duan et al, 2018 or masks (Rapson et al, 2018).…”

“…The provident detection of objects was already studied by other authors: In daylight, Naser, 2019 providently detected objects by analyzing shadow movements; At nighttime, Oldenziel et al, 2020 andSaralajew et al, 2021 studied the task to providently detect oncoming vehicles by detecting light artifacts produced by their headlights. Oldenziel et al, 2020 analyzed the discrepancy between the human abilities and an in-production computer vision system in detecting oncoming vehicles.…”

“…Together with the dataset published by Bell et al, 2021, these two datasets are the only large-scale datasets publicly available for the detection of vehicles at nighttime and annotated by keypoints-other available datasets for this task use bounding boxes (e. g., Rezaei & Klette, 2017, Duan et al, 2018 or masks (Rapson et al, 2018). By using keypoints for the annotation of light artifacts, Saralajew et al, 2021 presented an approach to reliably annotate light reflections, which are so fuzzy and weak in intensity that they cannot be objectively annotated by bounding boxes (as concluded by Oldenziel et al, 2020 and showed in a test by Saralajew et al, 2021). Similar to Bell et al, 2021, using the keypoints as initial seeds, the authors further explored methods to extend the keypoint annotations to bounding boxes with low annotation uncertainty so that state-of-the-art object detection methods can be applied.…”

“…1 The three characteristic states how an oncoming vehicle becomes visible. The images correspond to the scene 309 of the PVDN dataset (Saralajew et al, 2021). Each visualization shows a cropped area of the entire image (shown in the upper right corner).…”