Integration of Image-Based Object Identification and Distance Estimation Algorithm for Field Operational Test System of Self-driving Vehicles

Camera sensors are widely used as recognition sensors because they have the advantage of a lower unit price than other sensors of the same specification and are still able to collect data in a form that is similar to the information acquired by the human eye. An autonomous vehicle is controlled by accurate perception of the environment and the situational context. In this study, we designed an algorithm that performs image-based object identification and distance estimation using one-off deep learning. Eleven major objects in a road environment, such as vehicle license plates, pedestrians, and vehicles, were defined, and a learning dataset was constructed by collecting actual image data from driving. To verify that the results of artificial intelligence learning were suitable for the road environment, the algorithm was tested on image data from a driver's perspective and from a Closed-Circuit Television (CCTV) perspective that were collected from the Daegu Field Operational Test (FOT) section. The accuracy of the object identification algorithm on the verification dataset was 95.6% on average, but the average accuracy for the CCTV perspective was 92%, which is slightly lower than the result for driver's perspective of 98%. After the verification, the integrated algorithm was proven to be applicable to the self-driving vehicle platform and the self-driving control center being operated in the Daegu’s self-driving FOT section.