Hongfei Yu

Effective background modelling and subtraction approach for moving object detection

This study presents a hierarchical background modelling and subtraction approach for real-time detection of moving objects. At the first level, a novel pixel-wise background modelling method is proposed for coarse detection. The method can dynamically assign the optimal number of components for each pixel with the borrow–lend strategy. And a flexible learning rate which is variable and different for each component is presented to adapt to scene changes. Additionally, a new mechanism using a framework of finite state machine is introduced to maintain and update the background models. At the second level, in order to deal with sudden illumination changes, a block-wise foreground validation approach is adopted for refined detection. The authors compare the proposed approach with state-of-the-art methods and experimental results under various scenes demonstrate the robustness and effectiveness of the proposed approach.

Real-Time Speed Limit Sign Detection and Recognition from Image Sequences

Traffic sign, especially speed limit sign recognition is important in a driver assistance system. In this paper, a robust approach for real-time detection and recognition of speed limit sign is presented. It consists of two major steps: sign detection and sign recognition. In detection stage, Fast Radial Symmetry Transform is utilized to detect possible sign locations. Then the new method proposed, that is named “Largest Containing Circle” is used to segment characters region. In recognition stage, one fuzzy template matching method is applied to coarse recognize the sign number character, Furthermore, we conduct a similar character recognizer based on the local feature vector for fine recognition of the character. Experimental results in different conditions, including sunny, cloudy, foggy and rainy weather demonstrates that most speed limit signs can be correctly detected and recognized with a high accuracy and the average processing time is 15ms per frame on a standard PC.

Moving Object Detection Using an In-Vehicle Fish-Eye Camera

This paper proposes a robust rear-view camera based moving object detection algorithm for backup aid and parking assist applications. A single fish-eye camera is used in order to get much larger FOV (field of view) for object detection. To detect various moving objects such as vehicles and pedestrians, the ego-motion of host vehicle is firstly estimated by A robust NGPP (near ground point projection) method. Then a novel point based moving object detection method is proposed which can detect fast motion as well as slight motion in the fish-eye image. Finally, a region based motion compensation method is used in order to filter out the false detection results caused by the error matching points. Experimental results under various conditions show that most of moving objects can be detected correctly by our algorithm.

Obstacle detection based on multiple cues fusion from monocular camera

This paper presents an obstacle detection method by using a monocular camera on a moving platform. The method can detect various static and moving obstacles based on multi-cues fusion. Firstly, an improved motion compensation cue is applied to detect the obstacles which violate the road plane assumption; Secondly, a novel image segmentation cue is introduced to increase the obstacle detection rate and decrease false detection rate; Furthermore, these cues are combined in a Bayesian framework to generate a probability map of obstacle; Finally, an efficient probability update model is presented to update the probability map and the obstacle regions can be generated. Experiment results show that the proposed obstacle detection method is robust to obstacle types, varying illumination conditions, and various scenes. Moreover, using the combined cues outperforms any individual cue.

A novel motion detection approach for large FOV cameras

Moving objects detection by an also moving camera plays an important role in driver assistance systems and robot navigations. Many motion detection methods have been proposed until now. But most of them are based on normal cameras with a limited view and not suitable for large FOV (field of view) cameras. For motion detection using large FOV cameras, there are two main challenges. One comes from difficulties to tell moving objects from moving background due to camera motion. The other comes from the image distortion brought by large FOV cameras. These two problems are solved in our approach by a novel motion detector which can be considered as a special motion constraint based on virtual planes. The experimental results under various scenes illustrate the effectiveness of this work.

Learning Based Combining Different Features for Medical Image Retrieval

In this paper, authors propose a new learning based method for medical image retrieval which is based on fusing different features by linearly combining different similarities. Considering the abundant classes of medical images, this paper avoid to train a classifier for each class by using large amount training data. Instead, by using optimization method to combine different features’ similarity, new method can get good performance while has no much training computation. Experimental results show that the algorithm has potential practical values for clinical routine application.