Deep Learning based Driver Assistance System (DL-DAS)

Paper Abstract:

Nowadays, cars are an indispensable part of our lives in transportation and people use them almost everywhere to transport. However, despite having such an important place, traffic accidents continue to occur all over the world. In this paper, the solution was found for the problem of traffic accidents with the "deep learning" models and "image processing" algorithms without any sensor. For that aim, images are categorized with %99.92 correctness into two classes which are “Night” and “Day” by a filter design. According to the results, algorithm follows a deep learning model selection which has %87.66 correctness for vehicles, and %80.47 correctness for people. Furthermore, for lane detection algorithm which has %88.80 correctness. In the light of these results, interface returns feedback to the driver.

Algorithm Procedure:

Filter creation algorithm

Filter algorithm

Flowchart

flowchart TD
    A[Start] --> B[Taking Image]
    B --> C[Filter Algorithm]
    C --> D{Is image from night?}
    D --> |YES| E[Use Model Tuned with Oxford Robotcar Dataset]
    D --> |NO| F[Use Pretrained Yolo Model with COCO]
    E --> G[Object and Lane Detection]
    F --> G
    G --> H{Are there any object between lanes?}
    H --> |YES| I[Distance Calculation]
    H --> |NO| L{Is our car inside between lanes?}
    I --> J{Are there enough distance?}
    J --> |Yes| L
    J --> |NO| K[Warn the driver!]
    L --> |YES| B
    L --> |NO| K
    K --> B

Loading

Results:

Day and night image matrice values:

Before filter	After filter

||