Detection of road markings for advanced driver assistance

• Main Author: Suchitra Sathyanarayana
• Other Authors: Thambipillai Srikanthan
• Format: Theses and Dissertations
• Language: English
• Published: 2013
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; DRNTU::Engineering::Computer science and engineering::Hardware::Arithmetic and logic structures; DRNTU::Engineering::Computer science and engineering::Computer systems organization::Special-purpose and application-based systems
• Online Access: https://hdl.handle.net/10356/54995
• Institution: Nanyang Technological University

Automatic detection of road markings will enhance the capabilities of Advanced Driver Assistance Systems (ADAS) as road markings denote vital information pertaining to traffic safety and navigation. However, little work has been done in the area of vision-based detection of road markings in general, and existing literature was found to be largely confined to lane marking detection. In this thesis, algorithms and architectures have been proposed to identify road markings that are categorized into (a) basic linear markings (BLM) (b) complex linear markings (CLM) (c) arrow markings and (d) pedestrian markings. The identification of BLM such as dashed and solid lane markings has been tackled first. An efficient feature extraction process based on gradient angle histograms was introduced for shortlisting potential lane marking candidates in a block-based manner, resulting in customized edge maps (called Straight Line Edge Map or SLEM). A study on the relationship between the block-size and the quality of SLEM was conducted to identify appropriate block settings leading to successful extraction of lane markings despite noisy edge pixels such as tree shadows. The outlier removal step involves decomposing the SLEM into positive and negative edges based on intensity transition characteristics, which are then subjected to the Hough Transform (HT). The thickness criterion was also introduced during the detection of HT peaks to eliminate artifacts resembling lane markings. The proposed GAH based preprocessing approach is shown to reduce the number of HT computations by more than 50%, due to a notable reduction in the number of edge pixels and angle range for HT. It is shown that the proposed lane marking detection technique yields high detection accuracy, ranging from 98% to 99%, upon validation on an extensive dataset with more than 6,000 image frames representing different illumination, weather and complex road conditions. In order to distinguish solid and dashed single lane markings, a two-pronged technique based on continuity analysis of the lane markings in the spatial and temporal domains was proposed. It was shown that this method successfully distinguishes dashed and solid lane markings for all the test sequences considered in the dataset. The proposed BLM detection module can be readily configured to detect horizontal BLMs such as stop lines. A generic version of the lane marking detection algorithm was realized to detect BLMs for a given thickness. A novel parallel method for HT computations called Additive Hough transform (AHT) was proposed to drastically collapse the complexity of HT computations by replacing the trigonometric operations with simple additions. It relies on dividing the image into uniform blocks, and processing them in parallel, by exploiting the proposed additive property of HT. A study on how block size affects the compute efficiency was conducted to show that AHT reduces the total computation time by at least k2 times as compared to existing HT architectures for a k  k grid, translating to a speed up of at least 67 times against existing methods for an 8  8 grid. It was demonstrated that the block size in AHT and SLEM can be adjusted to achieve the required speedup and improve the quality of edge map, respectively. Detecting CLMs, such as double linear markings and zigzag markings, is more challenging than detecting BLM, particularly due to inherent limitations of the conventional Hough Peak detection. This motivated the development of the novel Adaptive Hough Peak Detection (AHPD) algorithm that adaptively identifies valid peaks in a noisy Hough space. The method relies on an iterative climbing process to generate a reduced Hough space, after which shape properties of peaks are exploited to isolate valid peaks from local minima. This not only eliminates redundant peak detections, but also detects lines in close proximity. The AHPD is also aided by the signed edges for separating lines in close proximity. Unlike conventional methods, AHPD is fully automatic and does not necessitate manual setting of the threshold and non-maximal suppression window size. This also makes it possible to automatically detect uneven lengths of lines as in the case of zigzag markings more deterministically. Evaluation of the AHPD on the 101 CMU object database confirms that it can successfully detect the major line features of interest, including lines of close proximity and lines of varying lengths amidst noise. Evaluation using 640 road images consisting of 120 zigzag images and 120 double linear lane markings shows that CLMs can be detected with 98% accuracy. Detection of road arrow markings is complex mainly due to varied appearances of the same marking under different scales and perspectives. A scale-invariant technique, based on shape signatures, is proposed for the efficient detection and classification of arrow markings. It relies on decomposing the arrow markings into its constituent parts by representing each part with a unique shape signature using a combination of GAH, signed edges and HT. The shape signatures associated with an arrow marking are not only invariant across scales and distortion effects, but are also easy to identify by employing methods proposed for BLM and CLM. This has led to the elimination of exhaustive matching operations associated with conventional template matching approaches. The method was evaluated on a database of over 200 arrow markings and it was shown to yield a detection and classification accuracy of 97% and 98% respectively. An efficient method to identify arrows in neighboring lanes was also proposed by relying on position and type of arrow detected in the host lane. It was shown to withstand the notable distortions and compression of the arrow marking features in the neighboring lanes. An algorithm to detect pedestrian markings was proposed next by incorporating techniques for detecting scale and distortion invariant shape signatures that are associated with the stripes of the pedestrian marking. Other road markings in the vicinity of the pedestrian crossing have also been employed to act as cues to predict an upcoming pedestrian marking so as to further enhance the robustness of the detection process. A unified engine to detect BLMs, CLMs, arrow and pedestrian markings is presented using the core modules such as GAH, signed edge generator, AHT and AHPD. Finally, the major contributions in this thesis have paved the way for propelling further research in this emerging area of interest and for realizing high performance ADAS at low cost.