The above article is about a survey on vehicle detection and tracking techniques completed by M. Swathy, P.S Nirmala and P.C. Geethu. They have proposed many methods and algorithms to detect vehicles on highways, based on information provided by video cameras and image processing techniques to captured and tested video sequences. They also worked within the project of designing a traffic management application and planning to extract exact traffic information by traffic image analyzing and controlling traffic flow by using vehicle count, vehicle velocity, traffic lane changes, vehicle trajectory, vehicle classification and vehicle density by tracking those vehicles. Their purpose is to provide real-time numerical data on traffic activity and to signal potentially abnormal situations through this project. They specially indicate that intelligent image detection systems are a most important approach to modern day traffic management. It is also important to most cost-effective and efficient traffic monitoring. They allowed that the automatic pre-processing for the operators to pick cameras to view and collect statistics with the aim to improve traffic flow. They focused on video cameras for a long time for traffic monitoring purposes, because they provide a rich information for human understanding.
They have highlighted different image processing techniques that can detect and identify vehicles for better traffic surveillance, using various tasks like foreground segmentation, feature extraction, background subtraction and threshold techniques.
There are four main vehicle detection methods they considered. Those are the object based approach, background subtraction method, feature-based methods and motion-based methods.
The object-based method has three steps such as segmentation, training and validation. For the segmentation, the process needed some training images from the total number of images depending upon the number of frames. In this model, the condition of applicable to high spatial resolution weakly sensed data, and to address the essential for a quantitative, user-supervised method for taking best segmentation parameters. It developed an impartial metric which is the number of training object matched with maximize area matched and is minimizes below and over segmentation for chosen images in objective primitives.
In background subtraction method also known as foreground, detection is a technique in the fields of image processing. The foreground is extracted for extra processing. The researchers pointed that the background subtraction can be varied, because of the intensity values of the background pixels. Those are done using non-parametric Kernel Density Estimation, Mixture of Gaussian's Uni-modal distributions and Adaptive multi-cue Background Subtraction. In background subtraction algorithm, the foreground vehicles are separated from the background and form a foreground mask. The threshold value concept is used in here. If the threshold value is less than the difference image, then it is taken as a moving object or otherwise taken as a background image. In fact that it can be completed using Gaussian Mixture Model (GMM), Adaptive multi-cue background subtraction, Averaging or Kalman filter.
The Gaussian Mixture Model is sampled and expected as diagonal, spherical, tied and full co-variance matrices supported from data. It provides the number of components properly. An adaptive multi-cue background subtraction obtains correct background subtraction result by using different hints. In the background averaging method, all video frames are added up. The learning rate requires the weight of a new frame and the background. This algorithm has little computational cost, however, it is likely to produce the tails behind moving objects due to contamination of the background with the appearance of the moving objects. Use the prompt background, which is the current frame with detected objects removed. The regions of detected objects are filled with the old background pattern. By averaging the instantaneous background, the tails generated by moving objects are reduced. The feedback of the motion mask could lead to erroneous background estimations if the threshold is set poorly. A dynamic threshold is applied to reduce this problem of never updating a region detected as foreground. A Kalman filter is used to estimate the background image, in this, the colour of each pixel is modelled by one filter.
In the feature-based method, the moving objects segmented from background image by collecting the set of features from the movement between the subsequent frames. This approach is based on learning which employs a set of labelled training data used for labelling the extracted objects features.
In motion-based method is about how much each image pixel moves between adjacent images. When it comes to the vehicle tracking they have stated five types of object tracking methods.
- Region-based tracking method
- Contour tracking method
- 3D model – based tracking method
- Feature-based tracking method
- Color and pattern-based method
Region-based tracking method is about to identifying a connected region in the image. This process is initialized by the background subtraction technique. A Kalman filter-based adaptive background model makes the background estimate to evolve as the weather and time of day might affect the lighting conditions. According to that research, this is suitable for free-flowing traffic. Region-based representation is reduced computational complexity.
In contour tracking method the basic idea is to make a representation of the bounding contour of the object and keep on dynamically updating it.
Model-based tracking method using 3D solid cuboid form.
Feature-based tracking method is an interactive and different framework. They stated that the proposed framework showed a good performance for the vehicle classification in surveillance videos regardless of significant challenges such as quality, limited image size, and large intra-class dissimilarities.
In colour and pattern-based method is used the technique by encompassed the YCrCb colour space for the construction preliminary background, vehicle location, vehicle tracking, shade elimination, segmenting foreground, and background updating algorithms. Through the experiments, this system proves that it works under several weather conditions, and is insensitive to lighting. A model-based system for traffic supervision visual tracking and classification of vehicles for multi-lane highway scene.
No comments:
Post a Comment