Visual Computing-based Perception System for Small Autonomous Vehicles: Development on a Lighter Computing Platform

Abstract

Recently, perception system for autonomous vehicle has seen a tremendous growth. Most of the recent works employ sensor fusion with complementary properties to produce a robust and accurate perceptive system for vehicle. However, this comes at a high price, requires high computing power and consumes more energy. In this study a perceptive system is designed to tackle the above issues while maintaining its accuracy and robustness. The proposed perceptive system is using only a pair of vision sensors. A Convolution Neural Network is used to detect and identify objects in the field of vision. A pair of cameras are then used to form a stereovision which is used to measure the distance of the objects detected. A disparity map from stereovision images was constructed first, then from the region of interest, a single disparity value was extracted to calculate the distance. The system is employed on a single board computer system StereoPi with the help of Intel Neural Compute Stick 2 to run deep neural network inference. An experiment was then conducted to test the perceptive system’s robustness, accuracy, and runtime. Results show that the proposed system is capable of a detection accuracy of 71.7% with an average error of 0.37% up to a distance of 1.3m.