Object tracking has long been an active research topic in image processing and computer vision fields with various application areas. For practical applications, the object tracking technique should be not only accurate but also fast in a real-time streaming condition. Recently, deep feature-based trackers have been proposed to achieve a higher accuracy, but those are not suitable for real-time tracking because of an extremely slow processing speed. The slow speed is a major factor to degrade tracking accuracy under a real-time streaming condition since the processing delay forces skipping frames. To increase the tracking accuracy with preserving the processing speed, this paper presents an improved kernelized correlation filter (KCF)-based tracking method that integrates three functional modules: (i) tracking failure detection, (ii) re-tracking using multiple search windows, and (iii) motion vector analysis to decide a preferred search window. Under a real-time streaming condition, the proposed method yields better results than the original KCF in the sense of tracking accuracy, and when a target has a very large movement, the proposed method outperforms a deep learning-based tracker, such as multi-domain convolutional neural network (MDNet).

In spite of excellent performance of deep learning-based computer vision algorithms, they are not suitable for real-time surveillance to detect abnormal behavior because of very high computational complexity. In this paper, we propose a real-time surveillance system for abnormal behavior analysis in a closed-circuit television (CCTV) environment by constructing an algorithm and system optimized for a CCTV environment. The proposed method combines pedestrian detection and tracking to extract pedestrian information in real-time, and detects abnormal behaviors such as intrusion, loitering, fall-down, and violence. To analyze an abnormal behavior, it first determines intrusion/loitering through the coordinates of an object and then determines fall-down/violence based on the behavior pattern of the object. The performance of the proposed method is evaluated using an intelligent CCTV data set distributed by Korea Internet and Security Agency (KISA).

Despite the significant benefits of low cost and scalability associated with monocular 3D object detection, accurately estimating depth from a single 2D image remains challenging due to the typical ill-posed nature of the problem. To address this issue, we propose a new method that improves depth estimation accuracy by randomly sampling object-wise points instead of relying on a single center point, which is a common practice in conventional methods. To generate the object-wise multiple reference points, we create a sampling space and obtain the ground truth by moving them from the sampling space to the object space. For this reason, the proposed approach is named ODD-M3D, which stands for Object-wise Dense Depth estimation for Monocular 3D object detection. In addition, we conduct an ablation study comparing LiDAR-guided and random sampling methods to identify the limitations of using point cloud data for image-based 3D object detection tasks. The proposed network achieved better performance by allowing for dense depth estimation instead of sparse depth estimation, which is typical in conventional networks.

In this paper, we present a coarse-to-fine convolutional neural network (CF-CNN) for learning multilabel classes. The basis of the proposed CF-CNN is a disjoint grouping method that first creates a class group with hierarchical association, and then assigns a new label to a class belonging to each group so that each class acquires multiple labels. CF-CNN consists of one main network and two subnetworks. Each subnetwork performs coarse prediction using the group labels created by the disjoint grouping method. The main network includes a refine convolution layer and performs fine prediction to fuse the feature maps acquired from the subnetwork. The generated class set in the upper level has the same classification boundary to that in the lower level. Since the classes belonging to the upper level label are classified with a higher priority, parameter optimization becomes easier. In experimental results, the proposed method is applied to various classification tasks to show a higher classification accuracy by up to 3% with a much smaller number of parameters without modification of the baseline model.

Abandoned object detection has recently been studied to detect dangerous situations such as illegally dumping explosives. Existing abandoned object detection used background subtraction, which is sensitive to noise and clutters of a similar size in outdoor environments. Unlike existing approaches, this article presents a deep learning-based detection method that is suitable for use in outdoor environments because it is robust to ghost effects and illumination changes. The proposed method consists of three parts: 1) pedestrian detection and tracking; 2) hand luggage detection (HLD); and 3) abandoned object decision. Since the proposed method detects the abandoned object through HLD instead of directly detecting the abandoned object itself, we could successfully improve the detection accuracy and robustness. The proposed method can also detect dangerous weapons or restricted hand luggage through HLD. This article presents an intelligent abandoned object detection system that can be applied for consumer applications.

Recently, the stereo imaging-based image enhancement approach has attracted increasing attention in the field of video analysis. This paper presents a dual camera-based stereo image defogging algorithm. Optical flow is first estimated from the stereo foggy image pair, and the initial disparity map is generated from the estimated optical flow. Next, an initial transmission map is generated using the initial disparity map. Atmospheric light is then estimated using the color line theory. The defogged result is finally reconstructed using the estimated transmission map and atmospheric light. The proposed method can refine the transmission map iteratively. Experimental results show that the proposed method can successfully remove fog without color distortion. The proposed method can be used as a pre-processing step for an outdoor video analysis system and a high-end smartphone with a dual camera system.