Ikuko Shimizu

Head pose determination from one image using a generic model

We present a new method for determining the pose of a human head from its 2D image. It does not use any artificial markers put on a face. The basic idea is to use a generic model of a human head, which accounts for variation in shape and facial expression. Particularly, a set of 3D curves are used to model the contours of eyes, lips and eyebrows. A technique called iterative closest curve matching (ICC) is proposed, which aims at recovering the pose by iteratively minimizing the distances between the projected model curves and their closest image curves. Because curves contain richer information (such as curvature and length) than points, ICC is both more robust and more efficient than the well-known iterative closest point matching techniques (ICP). Furthermore, the image can be taken by a camera with unknown internal parameters, which can be recovered by our technique thanks to the 3D model. Preliminary experiments show that the proposed technique is promising and that an accurate pose estimate can be obtained from just one image with a generic head model.

Optimal Consensus set for digital line and plane fitting

This article presents a new method for fitting a digital line or plane to a given set of points in a 2D or 3D image in the presence of noise by maximizing the number of inliers, namely the consensus set. By using a digital model instead of a continuous one, we show that we can generate all possible consensus sets for model fitting. We present a deterministic algorithm that efficiently searches the optimal solution with time complexity O(Nd log N) for dimension d, where d = 2,3, together with space complexity O(N) where N is the number of points.

Sensor fusion-based pedestrian collision warning system with crosswalk detection

This paper describes a pedestrian collision warning system with crosswalk detection feature based on sensor fusion of a monocular camera and a millimeter wave radar. The method to decide about the presence of a pedestrian is based on the assumption that objects moving along a crosswalk can be interpreted as pedestrians under certain circumstances. The advantage of the described solution is its robustness and effectiveness since it is limited to crosswalks. The camera can be used to detect the crosswalk. Data from both sensors can then be used to infer about the presence of a pedestrian. The sensor fusion algorithm which combines data from the sensors is explained. Then the paper describes a warning concept which provides auditory alarm and visual information about the presence of a crosswalk as well as pedestrians to a driver, depending on the estimated collision probability. Finally, test drives on an experimental vehicle are presented and the results verify that the proposed warning system is running in practice.

Digital planar surface segmentation using local geometric patterns

This paper presents a hybrid two-step method for segmenting a 3D grid-point cloud into planar surfaces by using discrete-geometry results. Digital planes contain a finite number of local geometric patterns (LGPs). Such a LGP possesses a set of normal vectors. By using LGP properties, we first reject non-linear points from a point cloud (edgebased step), and then classify non-rejected points whose LGPs have common normal vectors into a planar-surface-point set (region-based step).

Optimal consensus set for digital plane fitting

This paper presents a method for fitting a digital plane to a given set of points in a 3D image in the presence of outliers. We present a new method that uses a digital plane model rather than the conventional continuous model. We show that such a digital model allows us to efficiently examine all possible consensus sets and to guarantee the solution optimality and exactness. Our algorithm has a time complexity O(N3 logN) together with a space complexity O(N) where N is the number of points.

Development of an On-Board Pedestrian Detection System Using Monocular Camera for Driver Assistance Applications

Recent statistical data of traffic accidents reveals that pedestrian fatalities are a particular priority in Europe and Japan. Many active safety systems for pedestrian protection based on sensor fusion approach such as millimeter wave radar, Laser Imaging Detecting And Ranging (LIDAR) and monocular/stereo camera are proposed. Aiming at large-scale system penetration in automobile markets, low-cost driver assistance system development becomes an important issue, therefore a monocular camera is one of solutions. To realize zero-traffic accident society, the objective of this paper is to develop a driver assistance system for pedestrian collision prevention based on using a monocular camera. As the concept of the system, moving objects on and nearby a crosswalk are interpreted as pedestrians. An on-board camera image processing algorithm is designed to detect the existence of a crosswalk in front of the vehicle. The feature extraction technique used in the algorithm is based on the feature called “cross ratio” of the crosswalk edges and the periodicity of the crosswalk paints. Then, nearby the detected crosswalk position, the region of interest is determined to be used in the moving object detection module. The position and the velocity of the moving object are obtained with the application of optical flow algorithm. Crosswalk image database in real-world traffic in Japan is constructed, and the precision of crosswalk detection is examined by using the database. Optical flow algorithm is applied on the region nearby the detected crosswalk in order to detect moving objects which are inferred as pedestrians. Image-based egomotion estimation is used to compensate the error in distance estimation and pedestrian movement. The effectiveness of the proposed system is verified by test drives. The system can perform the detection of crosswalks in urban area in various weather conditions with high detection rate. Pedestrians on crosswalks can also be detected by Optical Flow-based image processing algorithm. The detectable range of the proposed pedestrian detection with crosswalk detection function is 20–25 m in front of vehicle. This approach does not claim to cover 100 % of all pedestrian accidents, but has the advantage of high robustness, low false alarm rate and cost efficient implementation. The feasibility of the proposed camera-based pedestrian detection system is shown in the paper. The validation of the crosswalk detection and pedestrian detection algorithm using real-world driving database will be conducted and demonstrated in the full paper.

A Linear Classification Method in a Very High Dimensional Space Using Distributed Representation

We have proposed a fast learning and classification method by using distributed representation of vectors. In this paper, first, we shows that our method provides faster and better performance than 1-NN method by introducing a definition of a similarity concerned with LSH scheme. Next we compare our method with the Naive Bayes with respect to the number of dimensions of features. While the Naive Bayes requires a considerably large dimensional feature space, our method achieves higher performance even where the number of dimensions of a feature space of our method is much smaller than that of Naive Bayes. We explain our method by formalizing as a linear classifier in a very high dimensional space and show it is a special case of Naive Bayes model. Experimental results show that our method provides superior classification rates with small time complexity of learning and classification and is applicable to large data set.

Automatic range image registration using mixed integer linear programming

A coarse registration method using Mixed Integer Linear Programming (MILP) is described that finds global optimal registration parameter values that are independent of the values of invariant features. We formulate the range image registration problem using MILP. Our algorithm using MILP formulation finds the best balanced optimal registration for robustly aligning two range images with the best balanced accuracy. It adjusts the error tolerance automatically in accordance with the accuracy of the given range image data. Experimental results show that this method of coarse registration is highly effective.

Graph-based range image registration combining geometric and photometric features

We propose a coarse registration method of range images using both geometric and photometric features. The framework of existing methods using multiple features first defines a single similarity distance summing up each feature based evaluations, and then minimizes the distance between range images for registration. In contrast, we formulate registration as a graph-based optimization problem, where we independently evaluate geometric feature and photometric feature and consider only the order of point-to-point matching quality. We then find as large consistent matching as possible in the sense of the matching-quality order. This is solved as one global combinatorial optimization problem. Our method thus does not require any good initial estimation and, at the same time, guarantees that the global solution is achieved.

Development of an On-Board Crosswalk Detection for Pedestrian Protection Using a Monocular Camera

Recent, there are many pedestrian accidents in Europe and Japan, and the rate of pedestrian fatalities is higher than the other accidents. These pedestrian accidents usually occur in complex environment such as residential or urban areas where many people exist. Many proposed active safety systems use the sensor fusion approach such as a millimeter wave radar, a LIDAR (Laser Imaging Detecting and Ranging) sensor and a monocular/stereo camera. However, it is difficult for these systems to prevent pedestrian accidents in complex environment. In addition, it is important for such active safety systems to be commercialized at a low cost and be equipped in a simple way, aiming at large-scale system penetration in automobile markets. Therefore, we developed an on-board crosswalk detection system by using a monocular camera and a warning system depending on pedestrian recognition. Crosswalk is located at the position where many people appear and the majority of pedestrian accidents occur. Based on the fact, the proposed system can achieve less false-alarm-rate of pedestrian detection by a combination of crosswalk detection and pedestrian detection. This paper shows the necessity of the crosswalk detection and warning system by the statistical data about pedestrian accidents near crosswalk. An experimental vehicle sensor and equipments used for testing are described. Next, this paper proposes an algorithm of the crosswalk detection and warning system including pedestrian detection with optical flow approach. Finally, this paper shows the feasibility of the system from the relationship between the detection distance and the rate of detection by using many video data captured in city road in Tokyo.