In this study the authors have presented a comprehensive approach to detection, tracking and recognition of traffic signs from a moving vehicle. Their system is comprised of three components. The detector utilizes a state-of-the-art object detection technique, but features a confidence-weighted mean shift mode-finding algorithm to improve its accuracy and cope with multiple redundant hypotheses in the detector's response space. The main contribution of their work are the novel tracking and recognition algorithms. The proposed tracker models the motion of the target through an instancespecific tracking function. It encodes correlations between the unique feature representation of a candidate sign and the affine distortions it is subject to while being approached by the camera. Based on the Lie group theory such a tracking function can be learned and updated instantly from random transformations applied to the image of the target in known pose. A detected and tracked sign candidate is classified by maximizing its similarity to the class's prototype image. This similarity is estimated by a linear combination of local image descriptor differences and is learned from image pairs using a novel variant of AdaBoost algorithm, called SimBoost. The proposed algorithms have been evaluated in a number of experiments involving static road sign images, synthetic image sequences, and real-life video captured with a car-mounted camera. The first experiment was aimed at evaluation of the detection refinement algorithm with two different object detection techniques and identifying the best-performing refined detector. The second experiment was intended to demonstrate the ability of the tracker to model the affine motion of the signs and reconstruct their frontal views under significant viewpoint changes. In the third experiment, they estimated the error rate of a classifier trained with different low-level image descriptors using the SimBoost algorithm. Based on the comparison of the obtained classification rates, they determined the most discriminative feature representation of the traffic signs. The overall performance of the system was measured based on the prototype C++ implementation and using realistic traffic video. The obtained results prove the efficiency of the presented algorithms and show that their approach could have good prospects for application on board of intelligent vehicles.
In-vehicle camera traffic sign detection and recognition
Verkehrzeichendetektion und -erkennung mit Bordkamera
Machine Vision and Applications ; 22 , 2 ; 359-375
2011
17 Seiten, 14 Bilder, 3 Tabellen, 28 Quellen
Article (Journal)
English
In-vehicle camera traffic sign detection and recognition
British Library Online Contents | 2011
|Traffic Sign Detection and Recognition for Intelligent Vehicle
British Library Conference Proceedings | 2011
|Traffic Sign Detection and Recognition
Springer Verlag | 2022
|Traffic Sign Detection and Recognition
Springer Verlag | 2017
|TRAFFIC SIGN RECOGNITION DEVICE AND TRAFFIC SIGN RECOGNITION METHOD
European Patent Office | 2023
|