Monodepth-based object detection and depth sensing for autonomous vehicle vision systems

Abstract

This study explores object recognition and distance measurement technologies for autonomous vehicles and addresses advancements in this area. Autonomous vehicles are vehicles capable of perceiving objects in their surroundings and moving safely without human intervention. The progress in these technologies holds the potential to revolutionize mobility, largely driven by the significant role of artificial intelligence in enabling these advancements. Artificial intelligence is considered a fundamental component for autonomous vehicles to perceive environmental conditions and ensure safe travel. Techniques such as deep learning and machine learning allow vehicles to process data collected through cameras, lidars, radars, and other sensors to interpret their surroundings and provide safe responses. The processes of object detection and depth estimation are continually being improved with traditional and AI-based methods.Deep learning models improve object detection and localization capabilities by processing complex image data, thereby enhancing the safety and performance of autonomous vehicles. Autonomous vehicles use various sensors to detect environmental objects and determine distances. These sensors include lidars, radars, cameras, and ultrasonic sensors. Data collected from these sensors are processed to perceive surrounding objects and determine distances from the vehicle. While camera sensors are widely used, distance estimation with single-lens 2D cameras can be challenging. Advancements in artificial intelligence techniques have achieved significant success with 2D camera images. Particularly, these techniques offer alternatives to challenges such as increasing sensor costs and the need for 3D cameras to adapt to various environmental conditions, enhancing the importance of progress in this area. In this study, different artificial neural network models were used for object detection and depth estimation. Optimization efforts were conducted to adapt trained models for real-world applications, and testing on the Nvidia Jetson device was considered a crucial step. The developed application achieved successful results in real-time processing performance, representing a significant advancement in autonomous vehicle technologies. The real-time processing performance of the developed application was determined to be 12 frames per second (FPS). Additionally, the average absolute error value obtained in the object detection and distance estimation areas was measured as 1.51 meters. These values indicate that the application operates quickly and reliably, accurately predicting object locations.