Please use this identifier to cite or link to this item:
http://hdl.handle.net/11375/26262
Title: | TEXT-BASED TRAFFIC SIGN DETECTION AND TRACKING IN VIDEO |
Authors: | Hu, Jianfeng |
Advisor: | Kirubarajan, Thiagalingam |
Department: | Electrical and Computer Engineering |
Publication Date: | 2021 |
Abstract: | In this thesis, the problem of multiple text-based traffic sign detection and tracking in video is explored, and the following specific problems are addressed: 1) localize the traffic sign in video by fusing the information from the front camera as well as other on-board data source such as rotation of the wheels and directional information of vehicles, 2) a CPU-based text-based traffic sign detector and directional parameter estimation for vehicles based on environmental conditions, 3) a text-based traffic signs detection and tracking framework for real-time application with a low cost data acquisition method. As a crucial component of Advanced Driver Assistance Systems (ADAS), traffic sign detection and tracking play essential roles in the Automotive Traffic Sign Detection and Recognition (ATSDR) system. Based on their different shapes, colors, letters and symbols, traffic signs in traffic scenario can be roughly categorized to two groups, namely, graphics-based (symbol-based) traffic signs and text-based traffic signs. Graphics-based traffic signs are regulatory signs, warning signs and temporary conditions signs, and text-based traffic signs are information and direction signs. Compared to graphics-based traffic signs, only a few algorithms have focused on text-based traffic signs detection and tracking. Detecting and tracking text-based traffic signs is a challenging task, mainly due to larger variations within the category and limited available dataset. Starting with localizing text-based traffic sign in urban traffic scenarios, the kinematic states of vehicles and the spatial-temporal relationships between vehicles and traffic signs need to be modelled and estimated. To solve this problem, a kinematic automotive motion model is proposed. This kinematic model fuses information from the front camera as well as other on-board data source such as rotation of the wheels and directional information of vehicles. Based on the proposed kinematic model, a text-based traffic sign localization algorithm is developed. The experiential results on real world video data show that the proposed localization algorithm achieves good performance and significantly reduces the computational cost compared to previously proposed methodologies. Next, the CPU-based text-based traffic sign detection method is studied. To relax the restriction of directional data acquisition in kinematic automotive motion model, a parameter estimation method is developed based on different environmental/weather conditions. Then, a search region definition approach for traffic signs detection is presented. This approach takes the advantages of spatial-temporal information of the previous frames and different kinematic vehicles motion models in video and largely reduces the massive and repeated detection for common Maximally Stable Extremal Regions (MSERs) detector. From the experiential results, the proposed approach achieves good performance in real-time applications. Finally, a text-based traffic sign detection and tracking framework is proposed for video-based Traffic Sign Recognition (TSR) system. In the detection stage, a data-driven text-based traffic signs detector is trained with street view images, and a low cost data acquisition approach is presented. In the tracking stage, a multi-traffic signs tracking algorithm is proposed based on kinematic automotive motion model. The framework is evaluated on both public Traffic Guide Panel dataset and our self-collected ETFLab Text-based Traffic Sign Video Dataset. The overall performance demonstrates the effectiveness of the proposed system, which can be better adapted to real-time applications. |
URI: | http://hdl.handle.net/11375/26262 |
Appears in Collections: | Open Access Dissertations and Theses |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
Hu_Jianfeng_202103_PhD.pdf | 35.41 MB | Adobe PDF | View/Open |
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.