Classifying Simulated Driving Scenarios from Automated Cars

Detection of driving scenarios is getting ever more importance for assessment and control of automated driving functions. This paper investigates the performance of two versions of a high-end 3D convolutional network for scenario classification. The first one uses fully 3D kernels, the second one separates, in each constituting block, the 2D spatial convolution from the temporal convolution, (2 + 1)D. We made the tests on a synthetic dataset created by specifying scenarios in OpenScenario and running them in the CarLA 3D simulator. We focused our analysis on three main performance profiles: at different frame per second rates, different video clip lengths, and different weather conditions. Results show an overall robustness of the 3D predictors, and seem to suggest two different use cases: (2 + 1)D looks more suited when the scenario changes quickly or a low latency is required, while the plain 3D solution is better for slow-changing scenarios and when FPS can be low.