Object Detection Research - Noah Borel

1. A Survey on Approaches of Object Detection (2013),  S. Shantaiya, K. Verma, and K. Mehta.

Overview:

- Object detection using video has progressed rapidly in recent years.

- In most cases the focus has been on human motion and behavior.

Introduction:

- Video detection is starting to be used in detection of pedestrians, monitoring car traffic and identifying strange behavior near ATMs.

- Video detection has a huge interest because of it's potential uses in security.

Object Detection Approaches:

- High quality but inexpensive video cameras as well as high-powered computers are used in object detection.

- The detection of moving objects in video is a key step in overall object detection.

Feature Based Object Detection:

- In this type of object detection the object is usually detected using shape, size or color.

Shape:

- Shape based object detection is very complex due to having to segment the object from the rest of the picture.

- It becomes increasingly harder when there are multiple objects in the picture with different shadings and light levels.

- A method for detection of athletes in game is the PCA-HOG. This involves transforming the athletes into Histograms of Oriented Gradient and then applying Principal Component Analysis to them.

- A limitation of this particular method is that it is difficult to detect multiple objects. 

- Another method is by using variable resolution double-level contours. This involves using a low resolution image to detect the object's edges and get the outline of the object. Then the high resolution image is used to in a process where the outline reduces until the actual object outline is found.

- One other method for shape based detection is to automatically differentiate between the foreground and background of the image and to detect the object based on the assumption that it has different geometrical features than the background.

Color:

- Color detection requires low computational power relative to other methods.

- A method for object detection using color involves the analyzing of RGB color images. A downside to this is that detection fails if the object is too small.

- Another color based detection involves using color histograms. This follows the idea of extracting HOG features with pixel colors inside being analyzed. A drawback of this method is that it can't work if the background is a similar color to the object.

Template Based Detection

- This requires a template of the object being detected to be available.

- Detection is done by matching the features of the template to the image being observed.

- The two types of template detection are fixed and deformable template matching.

- Fixed template matching is very useful if the object's shape does not vary with the change in viewing angles.

- Deformable template matching is more suitable in most cases of detection because objects tend to not have the same shape from different angles of view.

Motion Based Detection

- This method relies on changes over pixel or block levels. 

- The first stationary frame is used as reference background frame. The preceding frames with the object in them are then compared to the reference frame and detect the object. This method requires the object to be moving continuously and not be in the image before being detected.