Object Detection and Application in Autonomous Vehicle

Name: Minh Duc Dao
Major: Mathematics, Computer Science
Advisors: Rob Kelvey, Max Taylor

Video instance segmentation is a type of deep learning algorithm that enables autonomous vehicles to perceive and interpret real-world scenes in real time. Since video footage comprises multiple static frames, a fast image instance segmentation algorithm can be utilized for video segmentation. For every object in a given image or video frame, the instance segmentation model must identify all pixels belonging to the object and assign a category label to the object. This describes the location, shape, size, and type of obstacle for the autonomous vehicle.

We note that an instance segmentation algorithm is an object detection algorithm that can generate a pixel-wise mask for each object instance. As the first step toward understanding the instance segmentation algorithm, we discuss the building block of two object detection algorithms in this study. An object detection algorithm must generate a bounding box and classification label for each object in the image. The bounding box describes the location and size of the object, while the label describes the object’s type.

Lastly, we compare the performance of the pre-trained Mask R-CNN (an instance segmentation model) and the pre-trained YOLOv5 (an object detection model) in detecting objects that appear in a traffic scene (street view). Through the comparison, we learn that YOLOv5 is a better algorithm for autonomous vehicle applications because it is able to localize and identify obstacles more correctly. However, neither model is fully prepared for deployment in real-world scenarios, as their performance rating falls below 60%, with 100% being absolutely accurate and correct.

Posted in Comments Enabled, Independent Study, Symposium 2023 on April 14, 2023.