N. S. Guzhva, R. N. Sadekov. Neural Network Based Algorithms for Traffic Lights Identification in Multi-Camera Driver Assistance Systems

The paper considers the problem of traffic lights recognition (detection, filtering and map-matching) using successive images in active tram driver assistance systems equipped with multiple cameras with different focal lengths. The procedure of the problem solution is described in detail, from measurements (detections) formed at the neural network output for each of the cameras, and up to the results matching with a map. In contrast to other studies of this subject, the authors of this work use 3D measurements as the output data for the neural network, and unscented Kalman filter (UKF) for determining the position of the traffic lights; in addition, a new method for fusing the data from two cameras is applied. The efficiency of the proposed algorithms and its modification has been field-tested. The results of experiments have shown that the algorithm provides the accuracy of 76.19% and completeness of 97.46% when used in combination with the tram control system with two cameras.

Keywords: computer vision, traffic lights detection, traffic lights recognition, monocular camera, unmanned tram

Guzhva N.S, Sadekov R.N. Neural Network Based Algorithms for Traffic Lights Identification in Multi-Camera Driver Assistance Systems / Gyroscopy and Navigation, 2024, Vol. 15, No.3, pp. 31-43.