Takashi Machida

GPU & CPU cooperative accelerated pedestrian and vehicle detection

This paper presents a fast pedestrian and vehicle detection framework that integrates GPU (graphics processing unit) and CPU implementations. We employ the Histograms of Oriented Gradients (HoG) and the Feature Interaction Descriptor (FIND) as object descriptors. FIND describes the high-level properties of an objects appearance by computing pair-wise interactions of adjacent region-level features. We also employ the cascade approach in a sliding-window manner with multi-classifiers specialized for both the direction of a pedestrian and the distance of the pedestrian to a camera installed on a vehicle. Although our detection framework can detect pedestrians and vehicles in images, it wastes computational cost on calculating high-dimensional FIND features. Therefore, to realize real-time processing, we utilize the NVIDIA CUDA framework for pedestrian and vehicle detection. Three parallelization techniques implemented on both a GPU and a CPU are incorporated in our detection framework. As a result, our proposed implementation can perform detection more than 30 times faster than can a conventional implementation running on a CPU. For a 640×480 image, our parallel techniques attain a processing speed of 23.8 fps (42 [ms]) and detects both pedestrians and vehicles.

Pedestrian detection by scene dependent classifiers with generative learning

Recently, pedestrian detection from in-vehicle camera images is becoming an crucial technology for Intelligent Transportation Systems (ITS). However, it is difficult to detect pedestrians accurately in various scenes by obtaining training samples. To tackle this problem, we propose a method to construct scene dependent classifiers to improve the accuracy of pedestrian detection. The proposed method selects an appropriate classifier based on the scene information that is a category of appearance associated with location information. To construct scene dependent classifiers, the proposed method introduces generative learning for synthesizing scene dependent training samples. Experimental results showed that the detection accuracy of the proposed method outperformed the comparative method, and we confirmed that scene dependent classifiers improved the accuracy of pedestrian detection.

Evaluation of image processing algorithms on vehicle safety system based on free-viewpoint image rendering

Development of algorithms for vehicle safety systems, which support safety driving, takes a long period of time and a huge cost because it requires an evaluation stage where huge combinations of possible driving situations should be evaluated by using videos which are captured beforehand in real environments. In this paper, we address this problem by using free viewpoint images instead of the real images. More concretely, we generate free-viewpoint images from a combination of a 3D point cloud measured by laser scanners and an omni-directional image sequence acquired in a real environment. We basically rely on the 3D point cloud for geometrically correct virtual viewpoint images. In order to remove the holes caused by the unmeasured region of the 3D point cloud and to remove false silhouettes in surface reconstruction, we have developed a technique of free-viewpoint image generation that uses both a 3D point cloud and depth information extracted from images. In the experiments, we have evaluated our framework with a white line detection algorithm and experimental results have shown the applicability of free-viewpoint images for evaluation of algorithms.

Improvement of Dead Reckoning in Urban Areas Through Integration of Low-Cost Multisensors

This paper presents a method of accurate dead reckoning in urban areas using low-cost sensors. In the evolution of advanced driver assistance systems, the seamless and accurate positioning of the vehicle has become one of the most important tasks, and dead reckoning plays an important role. Visual odometry is one of the most attractive approaches for this dead reckoning, but in urban areas, the accuracy of visual odometry is degraded due to the surrounding moving objects. Moreover, the error of heading estimation used with the visual odometry cannot be corrected with satellite information, due to poor satellite signal reception. In this study, solutions of these problems are presented to improve the accuracy of the visual odometry in urban environments. The first key technique is moving object detection using inertial measurement unite (IMU) and pattern recognition, which improves the robustness of visual odometry in the dynamic environments. The second key technique is heading estimation using time-series tightly coupled integration of satellite Doppler shift and IMU, which makes heading correction possible where there is poor satellite reception. In evaluation experiments in urban areas, the error of dead reckoning using this proposed method is reduced to about one-fourth compared to the conventional approach.

Toward human-like lane following behavior in urban environment with a learning-based behavior-induction potential map

In order to achieve harmony in the mixed traffic, it is crucial to have autonomous vehicles behave like human drivers. This work addresses a vision-based approach toward human-like lane following behavior in complex urban environment. At first, a deep architecture is adopted to generate a set of vehicle hypotheses. Subsequently, a hybrid merging procedure is performed to jointly output the final detection results based on both the image evidence and the statistical support of vehicle hypotheses. After that, the detected vehicles are classified into six categories by Bayesian Network, i.e., leader vehicle, parking vehicle, tail-end vehicle, exiting vehicle, merging vehicle and other vehicle. With this information, a learning-based instance-level behavior-induction potential map is constructed to generate a safe as well as reasonable local path for following a predefined lane-level route. Experimental results in various typical but challenging urban traffic scenes substantiated the effectiveness of the proposed approach.

Moving object detection from a point cloud using photometric and depth consistencies

3D models of outdoor environments have been used for several applications such as a virtual earth system and a vision-based vehicle safety system. 3D data for constructing such 3D models are often measured by an on-vehicle system equipped with laser rangefinders, cameras, and GPS/IMU. However, 3D data of moving objects on streets lead to inaccurate 3D models when modeling outdoor environments. To solve this problem, this paper proposes a moving object detection method for point clouds by minimizing an energy function based on photometric and depth consistencies assuming that input data consist of synchronized point clouds, images, and camera poses from a single sequence captured with a moving on-vehicle system.