This paper presents a patch-based approach for crowd density estimation in public scenes. We formulate the problem of estimating density in a structured learning framework applied to random decision forests. Our approach learns the mapping between patch features and relative locations of all objects inside each patch, which contribute to generate the patch density map through Gaussian kernel density estimation. We build the forest in a coarse-to-fine manner with two split node layers, and further propose a crowdedness prior and an effective forest reduction method to improve the estimation accuracy and speed. Moreover, we introduce a semi-automatic training method to learn the estimator for a specific scene. We achieved state-of-the-art results on the public Mall dataset and UCSD dataset, and also proposed two potential applications in traffic counts and scene understanding with promising results.

While online approaches for multi-person tracking are still hard to deal with occlusions, many tracking methods use a global data association approach to find person trajectories on a long segment of consecutive frames. Such offline approaches often cause long output latencies, which are unsuitable for camera-based surveillance systems. Based on an observation that the current position of an object should be estimated from both its past and future positions that are observed in a short period of time, we propose a novel object tracking method based on a dynamic programming framework. This method iteratively associates and integrates track lets obtained by visual tracking which are observed just before and after occlusions. We always produce outputs at a constant short delay time, and requires only the standard HOG detector for high performance. Our tracking method can run at 10-60 fps on a single CPU core, and attains the state-of-the-art performance for the Town Center dataset.

This paper present a part-based approach for detecting objects with large variation of appearance. We extract local image patches as local features both from the object and from the background in training images to learn an object part model discriminatively. Our object part model discriminates the local features whether they are an object part or not. Based on the discrimination results, each local feature casts probabilistic votes for the object location and size which are learned from the training images. Our object part model also requires regression performance for predicting the object location and size through the voting procedure. We build such an object part model with an ensemble of randomized trees trained by splitting each tree node so as to reduce the entropy of class label distribution and the variance of object location and size. Experimental results on hand detection with large pose variation show that our approach outperforms conventional generalized Hough transform. We verified the performance on a public dataset of side-view cars.

We present a novel method for detecting vehicles as obstacles in various road scenes using a single onboard camera. Vehicles are detected by testing whether the motion of a set of three horizontal line segments, which are always on the vehicles, satisfies the motion constraint of the ground plane or that of the surface plane of the vehicles. The motion constraint of each plane is derived from the projective invariant combined with the vanishing line of the plane that is a prior knowledge of road scenes. The proposed method is implemented into a newly developed onboard LSI. Experimental results for real road scenes under various conditions show the effectiveness of the proposed method.

This paper describes an active camera system for high-speed object tracking. Under ordinary illumination, such as fluorescent lights, our system is capable of tracking an object in a typical indoor environment that contains a cluttered background. In order to accomplish this task, we developed a high-sensitivity high-frame-rate visual sensor system and a real-time object tracking algorithm that makes use of our previously proposed motion estimation technique, which is capable of robustly estimating optical flow in a high-frame-rate image sequence. The experimental results for our active camera system show the effectiveness and robustness of our visual sensor system and tracking algorithm.

We propose a robust method of detecting obstacles using a single camera mounted on a vehicle. In order to distinguish an obstacle from the ground plane, an effective criterion is to utilize motion parallax between them. We set two virtual planes corresponding to an obstacle and the ground plane respectively, track them simultaneously in the image without finding correspondences of feature points, and detect an obstacle by selecting the virtual plane corresponding to obstacles whose tracking result is more consistent with the motion of pixels than the one with respect to the ground plane. Employing image sequences captured at a high frame rate makes it possible to track virtual planes accurately and realize robust obstacle detection.

This paper describes a method of tracking a person with 3D translation and rotation by integrating optical flow and depth. The target region is first extracted based on the probability of each pixel belonging to the target person. The target state (3D position, posture, motion) is estimated based on the shape and the position of the target region in addition to optical flow and depth. Multiple target states are maintained when the image measurements give rise to ambiguities about the target state. Experimental results with real image sequences show the effectiveness of our method.