eTraM: Event-based Traffic Monitoring Dataset

eTraM Statistics

The eTraM dataset consists of 10 hours of data collected from the Prophesee EVK4 HD camera. In addition to the annotated static perception data, the dataset includes sequences of ego-motion event-based data, which provides increased diversity and opportunities for further experimentation.

The text presents an analysis of the average time spent by instances from different classes at traffic sites. The analysis reveals distinct temporal dynamics for the various classes in the dataset. Pedestrians and wheelchair users spend the most time at the traffic sites, reflecting their slower movement speeds. In contrast, vehicle classes tend to spend relatively less time at the sites.

The text analyzes the distribution of different categories (VH, PED, and MM) based on the area they cover - small, medium, and large, as shown in Figure 2.

The paper establishes benchmarks based on object size classifications, as shown in Table 1. Analysis reveals similar performance trends across the models. For pedestrian and vehicle categories, performance is consistently superior on medium-sized instances compared to small and large-sized. Vehicle performance is similar across all three size classifications, but pedestrian performance significantly drops with small-sized instances. In contrast, micro-mobility performs better on small-sized instances than medium-sized, although its best-performing size is still worse than the worst-performing of pedestrian and vehicle.

These results indicate a performance degradation when dealing with small-sized objects, particularly micro-mobility. This may be due to the lack of contour and color information in the raw event data. Figure 3 also presents the frequency of aspect ratios for each class in the eTraM dataset.

Denoising Using Spatiotemporal Filter

The provided text discusses a denoising step implemented for the eTraM system to address the noise present in the event stream, particularly during nighttime data with increased levels of reflections and pointed light sources from streets and vehicles. Figure 4 qualitatively illustrates the effectiveness of the spatiotemporal filter [3] by presenting a side-by-side comparison of images before and after applying the filter, showcasing the impact of noise reduction on event data frames.

Implementation Details

The paper examines how well event-based models perform on the eTraM dataset. The authors trained three state-of-the-art architectures - RVT, RED, and YOLOv8 - on 7 hours of data, and evaluated them on 1.5 hours of validation and test data. Different learning rates were used for the models.

The paper discusses two input representations used in the experiments: Histogram of Events and Time Surfaces. Histogram of Events assigns each event to a cell based on its position and timestamp, and tallies the counts in each cell and time bin, with separate counts for each polarity. Time Surfaces record the timestamp of the most recent event for each pixel, with an exponential decay applied to diminish the influence of older events. The input representations are used as three-dimensional tensors, with the dimensions corresponding to the relevant parameters.

The mathematical formulations for updating the Histogram of Events and Time Surfaces are provided in the paper.

Detection and Tracking Examples

The provided section discusses the detection results using tensor-based methods, RVT and RED, on the eTraM dataset. The detection results are used to perform tracking using an IoU-based thresholding technique, which yields a Multi-Object Tracking Accuracy (MOTA) of 0.18 and a Multi-Object Tracking Precision (MOTP) of 0.28 on the eTraM test set. The paper notes that the precise evaluation of tracking performance is made possible by the inclusion of object IDs within the eTraM dataset. An example of ground truth objects and their corresponding tracking is illustrated in Figure 7.