Initially, we recorded joystick keys (left, right, forward and backward) and driver view images. We trained simple feed forward network with four convolution and four fully connected layers on 10,000 instances. Due to small dataset with inherent noise of joystick key strokes, model could not perform well on test set. Therefore, we wrote a Mod to collect actual parameters vehicle from game environment.
The result of this model tested in real time in GTA-V is given video. Since, the dataset was small, and had binary command. Therefore, we explored GTA-V mods in C# to get information from game engine and written a mod to collect stable dataset. We collected a dataset containing around 45,000 samples where each instance has driver view image, vehicle steering angle, speed, breaking mode, engine torque etc., in total 8 parameters. We trained earlier mentioned network and ResNet34 to predict only two parameters; steering and speed. Then, we tested both models in real time in GTA-V. The video depict the results. Analysis:
0 Comments
Leave a Reply. |
Archives
August 2019
Categories
|