I7 8700 cpu

The segmentation architectures UNet and DeepLab v3 were selected as the training models. The UNet and DeepLab network architectures were each implemented at two depths: UNet18, UNet36, DeepLab50 and DeepLab101. These four AI models were then trained for the segmentation tasks. All the models were trained using the Adam optimisation algorithm at 200 epochs. The architecture was trained in the Pytorch framework (learning rate, 0.0001; batch size, 8) to achieve the best possible validation loss. All the processing work was conducted on a PC with an Intel i7-8700 CPU, 32GB RAM and a single Nvidia RTX 2080 Ti GPU with 12G VRAM (Jumbo computer supplies, Hong Kong, China). Figure 4 presents the network architecture of the UNet model.

The driving evaluation was conducted with an F1 simulator (The Codemasters Software Company Limited). The simulation was run on a PC, using Windows 10 (i7-8700 CPU @ 3.20 GHz, NVIDIA GeForce GTX 1070 graphics card). The visual system consisted of a 27-in screen (BenQ XL2720T), with a resolution of 1,920 × 1,080 pixels, a refresh rate of 120 Hz and a latency of 1 ms. Auditory feedback was provided through external speakers set to ∼60 dB for all participants. To pilot the simulation, we used a steering wheel (ClubSport wheel Formula Carbon, Fanatec^®, Landshut, Germany), throttle, and brake pedals (CSL Elite Pedale LC, Fanatec^®, Landshut, Germany). The steering wheel was mounted on wheel-base pedals (ClubSport Wheel Base V2.5, Fanatec^®, Landshut, Germany) providing force feedback. The driving aids traction control and ABS were fully enabled, while gear shifting was set to manual. All other driving aids were disabled. Participants were tested under the same conditions: track, Austria; car, Ferrari; perspective, third person; session, practice (without opponents).