Soon Kwon

Stereo vision–based vehicle detection using a road feature and disparity histogram

This paper presents stereo vision-based vehicle detection approach on the road using a road feature and disparity histogram. It is not easy to detect only vehicles robustly on the road in various traffic situations, for example, a nonflat road or a multiple-obstacle situation. This paper focuses on the improvement of vehicle detection performance in various real traffic situations. The approach consists of three steps, namely obstacle localization, obstacle segmentation, and vehicle verification. First, we extract a road feature from v-disparity maps binarized using the most frequent values in each row and column, and adopt the extracted road feature as an obstacle criterion in column detection. However, many obstacles still coexist in each localized obstacle area. Thus, we divide the localized obstacle area into multiple obstacles using a disparity histogram and remerge the divided obstacles using four criteria parameters, namely the obstacle size, distance, and angle between the divided obstacles, and the difference of disparity values. Finally, we verify the vehicles using a depth map and gray image to improve the performance. We verify the performance of our proposed method by conducting experiments in various real traffic situations. The average recall rate of vehicle detection is 95.5%.

Distance estimation algorithm for both long and short ranges based on stereo vision system

We present a distance measurement method based on stereo vision system while guaranteeing accuracy and reliability. It has been considered as difficult problem to measure both long and short distance with a stereo vision system accurately due to sampling error and camera sensor error. To resolve these problems of the stereo vision system, we utilize an algorithm which is consisted of a modified sub-pixel displacement method to enhance the accuracy of disparity and strong tracking Kalman filter (STKF) to reduce the camera sensor errors. Our displacement method and the usefulness of STKF are verified as compared to other displacement methods and conventional Kalman filter (CKF) through simulating on the several distance ranges. The Monte-Carlo simulation results show that our algorithm is capable of measuring up to hundreds of meters while root mean square error (RMSE) maintains about 0.04 at all ranges, even though the target vehicle maneuvers or moves nonlinearly.

Obstacle localization with a binarized v-disparity map using local maximum frequency values in stereo vision

In this paper, we propose an obstacle localization method using column detection with a binarized v-disparity map. For localizing obstacles robustly in environments where there exist many obstacles, such as roadside trees, pedestrians, or where median strips exist, we also propose a new method which extracts a road feature. We create a binarized v-disparity map using local maximum frequency values in each row for emphasizing a diagonal straight line, namely a road feature. And to further eliminate noise, we use a comparing method which compares all road feature values with median values. Finally, we use a linear interpolation for rows which have no value. We can extract a road feature through this method robustly. And we adopt this new standard to localize obstacles. An experimental result which uses a real road image proved that our proposed method has the advantage of extracting a road feature and localizing obstacles in environments where many obstacles exist.

A fusion method of data association and virtual detection for minimizing track loss and false track

In this paper, we present a method to track multiple moving vehicles using the global nearest neighborhood (GNN) data association (DA) based on 2D global position and virtual detection based on motion tracking. Unlikely the single target tracking, multiple target tracking needs to associate observation-to-track pairs. DA is a process to determine which measurements are used to update each track. We use the GNN data association not to lost track and not to connect incorrect measurements. GNN is a simple, robust, and optimal technique for intelligent vehicle applications with a stereo vision system that can reliably estimates the position of a vehicle. However, an incomplete detection and recognition technique bring low track maintenance due to missed detections and false alarms. A complementary virtual detection method adds to GNN method. Virtual detection is used to recover the missed detection by motion tracking when the track maintains for some periods. Motion tracking estimates virtual region of interest (ROI) of the missed detection using a pyramidal Lukas-Kanade feature tracker. Next, GNN associates the lost tracks and virtual measurements if the measurement exists in the validation gate. Our experimental results show that our tracking method works well in a stereo vision system with incomplete detection and recognition ability.

A sliced synchronous iteration architecture for real-time global stereo matching

In this paper, we present a low memory-cost message iteration architecture for a fast belief propagation(BP) algorithm. To meet the real-time goal, our architecture basically follows multi-scale BP method and truncated linear smoothness cost model. We observe that the message iteration process in BP requires a huge intermediate buffer to store four directional messages of the whole node. Therefore, instead of updating all the node messages in each iteration sequence, we propose that individual node could be completed iteration process in ahead and consecutively execute it node by node. The key ideas in this paper focus on both maximizing architectures parallelism and minimizing implementation cost overhead. Therefore, we first apply a pipelined architecture to each iteration stage that is executed independently. Note that pipelining makes it faster message throughput at a single iteration cycle rather than consuming whole iteration cycle time as previously. We also make multiple message update nodes as a minimal processing unit to maximize the parallelism. For the multi-scale BP method, the proposed parallel architecture does not cause additional execution time for processing the nodes in the down-scaled Markov Random Field(MRF). Considering VGA image size, 4 iterations per each scale and 64 disparity levels, our approach can reduce memory complexity by 99.7% and make it 340 times faster than the general multi-scale BP architecture.

Stereo vision-based improving cascade classifier learning for vehicle detection

In this article, we describe an improved method of vehicle detection. AdaBoost, a classifier trained by adaptive boosting and originally developed for face detection, has become popular among computer vision researchers for vehicle detection. Although it is the choice of many researchers in the intelligent vehicle field, it tends to yield many false-positive results because of the poor discernment of its simple features. It is also excessively slow to processing speed as the classifiers detection window usually searches the entire input image. We propose a solution that overcomes both these disadvantages. The stereo vision technique allows us to produce a depth map, providing information on the distances of objects. With that information, we can define a region of interest (RoI) and restrict the vehicle search to that region only. This method simultaneously blocks false-positive results and reduces the computing time for detection. Our experiments prove the superiority of the proposed method.

Stereo vision-based obstacle detection using dense disparity map

In this paper, we propose stereo vision-based obstacle detection method on the road using a dense disparity map. We use the dense disparity map to detect obstacles robustly in real traffic situations. Our method consists of three stages, namely road feature extraction, column detection, obstacle segmentation. First, we extract a road feature from a v- disparity map calculated from a dense disparity map. And we perform a column detection using the extracted road feature as a criterion that decides whether obstacles exist or not. Finally, we perform a segmentation using a birds-eye view mapping to divide the merged obstacle into each obstacle accurately. We conduct experiments to verify our method in the real traffic situations.

ResNet-Based Vehicle Classification and Localization in Traffic Surveillance Systems

In this paper, we present ResNet-based vehicle classification and localization methods using real traffic surveillance recordings. We utilize a MIOvision traffic dataset, which comprises 11 categories including a variety of vehicles, such as bicycle, bus, car, motorcycle, and so on. To improve the classification performance, we exploit a technique called joint fine-tuning (JF). In addition, we propose a dropping CNN (DropCNN) method to create a synergy effect with the JF. For the localization, we implement basic concepts of state-of-the-art region based detector combined with a backbone convolutional feature extractor using 50 and 101 layers of residual networks and ensemble them into a single model. Finally, we achieved the highest accuracy in both classification and localization tasks using the dataset among several state-of-the-art methods, including VGG16, AlexNet, and ResNet50 for the classification, and YOLO Faster R-CNN, and SSD for the localization reported on the website.

Various lane marking detection and classification for vision-based navigation system?

A vision-based car navigation system (CNS) gives drivers more precise and realistic traffic data than a traditional 2D-CNS. As part of the vision-based CNS, the ability to detect lane markings can provide significant warnings which increase traffic safety and convenience. Meanwhile, accurate lane classification results can indicate the current/approaching road conditions in this system. This paper concentrates on two kernels: lane marking detection and lane type identification. The lane detection part uses IPM and histogram sampling and the lane marking type classification step utilizes spatial and frequency sampling for different types of lane markings.

Three plot correlation-based small infrared target detection in densesun-glint environment for infrared search and track

This paper presents a separate spatio-temporal filter based small infrared target detection method to address the sea-based infrared search and track (IRST) problem in dense sun-glint environment. It is critical to detect small infrared targets such as sea-skimming missiles or asymmetric small ships for national defense. On the sea surface, sun-glint clutters degrade the detection performance. Furthermore, if we have to detect true targets using only three images with a low frame rate camera, then the problem is more difficult. We propose a novel three plot correlation filter and statistics based clutter reduction method to achieve robust small target detection rate in dense sun-glint environment. We validate the robust detection performance of the proposed method via real infrared test sequences including synthetic targets.