DefinePK

Research Objective

To develop a cost-effective visual perception model for self-driving cars that focuses on lane following using only camera input and Convolutional Neural Networks (CNNs).

Methodology

The study proposes a model that utilizes a monocular camera as the sole sensor for environmental perception. A Convolutional Neural Network (CNN), specifically AlexNet, was trained in a simulated environment (Audacity simulator) using 120x120 grayscale images. The model was trained on data collected directly from the simulator. The action space for the car was defined as [left, forward, right]. The trained model was then tested within the same simulated environment.

Methodology Flowchart

                        graph TD;
    A["Collect Data from Simulator"] --> B["Preprocess Images: 120x120 grayscale"];
    B --> C["Train AlexNet CNN Model"];
    C --> D["Model Achieves Target Accuracy?"];
    D -- Yes --> E["Test Model in Simulator"];
    D -- No --> C;
    E --> F["Predict Car Action: Left/Forward/Right"];
    F --> G["Execute Action in Simulator"];
    G --> H["Evaluate Performance"];                    

Discussion

The paper argues that relying solely on visual input, specifically camera data, is a more cost-effective approach compared to using expensive sensors like RADAR and LiDAR for self-driving car perception. The use of a simulated environment for training and testing simplifies the process and allows for efficient data collection and model validation. The AlexNet architecture was chosen for its established performance in image recognition tasks.

Key Findings

The trained CNN model achieved 96.75 percent accuracy with a 0.082 percent loss rate after approximately 13 hours of training. The model was capable of detecting road lanes, identifying the middle line, and predicting the appropriate action (left, forward, or right) for the car to follow the lane.

Conclusion

The presented visual perception deep drive model, trained using AlexNet CNN in a simulated environment, successfully demonstrated the ability to follow road lanes for a self-driving car. This approach offers a viable and cost-effective alternative to sensor-heavy systems.

Fact Check

1. Training Time: The model took approximately 13 hours to train. (Confirmed in "Model Training" section).
2. Accuracy: The model achieved 96.75 percent accuracy. (Confirmed in "Model Training" section).
3. Image Resolution: The input image resolution was reduced to 120x120 pixels. (Confirmed in "Preprocessing" section).

Mind Map