Deep Learning for Autonomous Vehicles

This group project attempted to predict the turning angle and speed of the autonomous vehicle. It was divided into two phases to outline the challenges mentioned above. The first part of the project was conducted through a Kaggle competition. In this phase, the team used a pre-existing dataset to train the model. This algorithm was then tested on a partially hidden test set to analyse its performance on unseen data. The second phase involved developing a model for a real car and assessing its performance on various tracks. The three tracks used for live-testing are given below:

Figure 8	T-Junction	Oval

After initially testing MobilNetV2, VGG16 and InceptionV3 the team moved towards a custom CNN model. This model consistently outperformed the transfer learning models while being far more efficient. Therefore, the team decided to use the custom CNN architecture as the basis for the primary model. The CNN splits into two parts from the input layer. Both paths are similar in architecture consisting of convolution, max-pooling, dense, and activation layers (as shown in the figure below).

Custom CNN Model Architecture

The size of the Convolution filter was 3X3 without any padding. On the contrary, The max pooling Kernel size was 2X2 (again, no padding). The activation layers used the ReLu activation function. In total, the model had 450,090 parameters. On the other hand, the output node for speed had binary cross-entropy as the loss function since this prediction task required binary classification. An extended version of stochastic gradient descent called the Adam optimiser was used to minimise the loss function. Additionally, a learning rate schedule was deployed on the model to stabilise this process. Class weighting was also used to balance the dataset for speed. L2 Regularisation was deployed to generalise this model. Dropout, which is a method that randomly ignores some layer outputs, was also used to further generalise the model. The dropout rate value used for this project was 0.5. The plot comparing the validation and training loss convergence is shown below.

Gradient Descent Plot for Validation and Training Sets

While the model performed well on the oval track, it struggled on the other tracks. Therefore, the team decided to collect more data for the T-junction and figure-8 tracks. Data augmentation was also applied to add more images of particular regions of the tracks where the model struggled. The final dataset consisted of around 27,347 images (the original dataset was around 13,500 images).