Applying CNN Models with Stereo Camera for Object Detection and Distance Estimation in Automatic Micro Electric Vehicle (MEVi)

Technological advancements in the automotive industry are currently focused on autonomous driving systems or driver assistance systems. Depth estimation is also an important feature of the autonomous driving system as it can accurately estimate the distance to the detected objects. In current developments technologies such as stereo cameras are more popular for a more accurate depth estimation. The process for object detection consists of a combination of classification, localization, and segmentation. Deep convolutional neural networks are the basis of object detection. The methods for object detection are divided into two categories namely the Region Proposal-based method such as RCNN, and the Classification-based method which provides for methods such as YOLO and MobileNet. As a part of the research conducted by the National Research and Innovation Agency (BRIN) on Automatic Micro Electric Vehicles or MEVi, this paper analyzed and compared the performance of several CNN models for object detection as well as the performance of depth sensing using ZED stereo camera which was used for the autonomous system in MEVi. Among several models that had been compared, YOLO v8 model demonstrates high accuracy and moderate speed as indicated by its high mAP (0.50) and FPS (38.46). Based on depth sensing results, the accuracy of distance measurement also depended on brightness and environment with the average of accuracy achieved indoors was 98.2% and outdoors was 96.3%.