A system for detecting objects and estimating their distance using a neural network

Abstract

This article proposes using neural networks to solve the challenge of accurately measuring the distance of an object using cameras and digital image processing. For this, a neural network was trained using a data set that includes information on the distance in pixels of the centers of mass of the object detected by the cameras. This data was used to teach the network to make an accurate estimate of the actual distance of the object. Image analysis methods were also used in conjunction with images of the object previously captured and trained with YoloV8 on Roboflow. The results obtained showed a notable improvement in the precision that is obtained when measuring the distance without the tedious calibration that is had in the other approaches considered for this investigation. Overcame the challenges associated with camera calibration due to possible distortion, accuracy, and generalization generated by changing the environment, resulting in an effective solution with 90% accuracy percentage and a dense neural network with an input layer, a hidden layer and an output layer with 2000 training cycles. These results demonstrate the potential of neural networks and image processing to address distance measurement problems in various applications, such as robotics, road safety, and autonomous navigation.