Andrew H. S. Lai

Lane detection by orientation and length discrimination

This paper describes a novel lane detection algorithm for visual traffic surveillance applications under the auspice of intelligent transportation systems. Traditional lane detection methods for vehicle navigation typically use spatial masks to isolate instantaneous lane information from on-vehicle camera images. When surveillance is concerned, complete lane and multiple lane information is essential for tracking vehicles and monitoring lane change frequency from overhead cameras, where traditional methods become inadequate. The algorithm presented in this paper extracts complete multiple lane information by utilizing prominent orientation and length features of lane markings and curb structures to discriminate against other minor features. Essentially, edges are first extracted from the background of a traffic sequence, then thinned and approximated by straight lines. From the resulting set of straight lines, orientation and length discriminations are carried out three-dimensionally with the aid of two-dimensional (2-D) to three-dimensional (3-D) coordinate transformation and K-means clustering. By doing so, edges with strong orientation and length affinity are retained and clustered, while short and isolated edges are eliminated. Overall, the merits of this algorithm are as follows. First, it works well under practical visual surveillance conditions. Second, using K-means for clustering offers a robust approach. Third, the algorithm is efficient as it only requires one image frame to determine the road center lines. Fourth, it computes multiple lane information simultaneously. Fifth, the center lines determined are accurate enough for the intended application.

Vehicle type classification from visual-based dimension estimation

This paper presents a visual-based dimension estimation method for vehicle type classification. Our method extracts moving vehicles from traffic image sequences and fits them with a simple deformable vehicle model. Using a set of coordination mapping functions derived from a calibrated camera model and relying on a shadow removal method, vehicles width, length and height are estimated. Our experimental tests show that the modeling method is effective and the estimation accuracy is sufficient for general vehicle type classification.

A fast and accurate scoreboard algorithm for estimating stationary backgrounds in an image sequence

This paper presents a stationary background estimation algorithm for color image sequences. The algorithm employs the running mode and running average algorithms, which are two commonly used algorithms, as the estimation core. A scoreboard is used to keep the pixel variations in the image sequence and is used to select between the running mode or the running average algorithm in each estimation. Our evaluation results show that by selecting, intelligently, the estimation core between the two algorithms according to the scoreboard values, the proposed background estimation algorithm has excellent performance in terms of estimation accuracy and speed.

Vehicle-type identification through automated virtual loop assignment and block-based direction-biased motion estimation

This paper presents a method of automated virtual loop assignment and direction-based motion estimation. The unique features of our approach are that: 1) a number of loops are automatically assigned to each lane. The merit of doing this is that it accommodates pan-tilt-zoom actions without needing further human interaction; 2) the size of the virtual loops is much smaller for estimation accuracy; and 3) the number of virtual loops per lane is large. The motion content of each block may be weighted and the collective result offers a more reliable and robust approach in motion estimation. Comparing this with traditional inductive loop detectors, there are a number of advantages. Our simulation results indicate that the proposed method is effective in type classification.

Effective moving cast shadow detection for monocular color traffic image sequences

For an accurate scene analysis using monocular color traffic image sequences, a robust segmentation of moving vehicles from the stationary background is generally required. However, the presence of moving cast shadow may lead to an inaccurate vehicle segmentation, and as a result, may lead to further erroneous scene analysis. We propose an effective method for the detection of moving cast shadow. By observing the characteristics of cast shadow in the luminance, chrominance, gradient density, and geometry domains, a combined probability map, called a shadow confidence score (SCS), is obtained. From the edge map of the input image, each edge pixel is examined to determine whether it belongs to the vehicle region based on its neighboring SCSs. The cast shadow is identified as those regions with high SCSs, which are outside the convex hull of the selected vehicle edge pixels. The proposed method is tested on 100 vehicle images taken under different lighting conditions (sunny and cloudy), viewing angles (roadside and overhead), vehicle sizes (small, medium, and large), and colors (similar to the road and not). The results indicate that an average error rate of around 14% is obtained while the lowest error rate is around 3% for large vehicles.

Detection of vehicle occlusion using a generalized deformable model

This paper presents a vehicle occlusion detection algorithm based on a generalized deformable model. A 3D solid cuboid model with up to six vertices is employed to fit any vehicle images, by varying the vertices for a best fit. The advantage of using such a model is that the number of parameterized vertices is small which can be easily deformed. Occlusion is detected by recording the changes in the Area Ratio and the dimensions of the generalized deformable model. Our tests show that the new modeling algorithm is effective in detecting vehicle occlusion.

Effective moving cast shadow detection for monocular color image sequences

For an accurate scene analysis in monocular image sequences, a robust segmentation of a moving object from the static background is generally required. However, the existence of moving cast shadow may lead to an inaccurate object segmentation, and as a result, lead to further erroneous scene analysis. An effective detection of moving cast shadow in monocular color image sequences is developed. Firstly, by realizing the various characteristics of shadow in luminance, chrominance, and gradient density, an indicator, called shadow confidence score, of the probability of the region classified as cast shadow is calculated. Secondly the canny edge detector is employed to detect edge pixels in the detected region. These pixels are then bounded by their convex hull, which estimates the position of the object. Lastly, by analyzing the shadow confidence score and the bounding hull, the cast shadow is identified as those regions outside the bounding hull and with high shadow confidence score. A number of typical outdoor scenes are evaluated and it is shown that our method can effectively detect the associated cast shadow from the object of interest.

Towards detection of moving cast shadows for visual traffic surveillance

In this paper, an effective method for the detection of moving cast shadow for visual traffic surveillance is proposed. Based on the cast shadow observations in luminance, chrominance, gradient density and geometry domains, a combined probability map, called the shadow confidence score, of the region belonging to the shadow is deduced. The object region is then separated from its shadow according to the score. The proposed method has been further tested on images taken under different lighting conditions (sunny and cloudy), viewing angles (roadside and overhead), and vehicle sizes (small, medium and large).

Performance evaluation of a feature‐preserving filtering algorithm for removing additive random noise in digital images

We evaluate the performance of a feature-preserving filtering algorithm over a range of images corrupted by typical additive random noise against three common spatial filter algorithms: median, sigma and averaging. The concept of the new algorithm is based on a corrupted- pixel identification methodology over a variable subimage size. Rather than processing every pixel indiscriminately in a digital image, this corrupted-pixel identification algorithm interrogates the image in variable-sized subimage regions to determine which are the corrupted pixels and which are not. As a result, only the corrupted pixels are being filtered, whereas the uncorrupted pixels are untouched. Extensive evalu- ation of the algorithm over a large number of noisy images shows that the corrupted-pixel identification algorithm exhibits three major charac- teristics. First, its ability in removing additive random noise is better vi- sually (subjective) and has the smallest mean-square errors (objective) in all cases compared with the median filter, averaging filter and sigma filter. Second, the effect of smoothing introduced by the new filter is minimal. In other words, most edge and line sharpness is preserved. Third, the corrupted-pixel identification algorithm is consistently faster than the median and sigma filters in all our test cases.

Modified CPI filter algorithm for removing salt-and-pepper noise in digital images

In this paper, the theoretical aspects, implementation issues, and performance analysis of a modified CPI filter algorithm are presented. As the concept of the original CPI algorithm is to identify corrupted pixels by interrogating subimages, and considering the intensity spread of pixel values within the subimage when making a decision, the modified algorithm similarly takes into account the subimage gray level distribution across the whole gray scale. It works on the assumption that to consider which group in the subimage is corrupted, the multiple- feature histogram representing a subimage gray level distribution must be transformed into a two-feature histogram such that these two features can be mapped onto the two available pixel classes. This transformation is performed by using a 1-sigma decision about the mean intensity of the subimage, which enables pixels that fall inside the sigma bounds to be considered as uncorrupted, and the rest corrupted. A performance analysis of the modified CPI, original CPI, average, median and sigma algorithms is given for noisy images corrupted by salt-and- pepper noise of the impulsive and Gaussian nature, and gray noise over the signal-to-noise ratios (SNR) of plus 50 dB to minus 50 dB. The results show that similar to the original CPI algorithm, the modified CPI algorithm exhibits a number of desirable features. Firstly, due to its pixel identification property, it has better noise removing capability than the conventional filter algorithms. Secondly, most features in the original image are preserved in the restored image compared with, say, the median filter. Thirdly, iterative filtering of a noisy image using the CPI algorithm is possible.