This section describes the background subtraction algorithms which can detect mov- ing objects. As seen in (Figure 2.4), background subtraction immediately follows the video capturing stage. Its objective is to extract the person’s silhouette from the video input by subtracting a background estimation model from this original
CHAPTER 2. IMAGE PROCESSING AND TRACKING METHODS
video input. This step allows the fall detection algorithm to determine features of the person that would seem important for fall detection stage. Before the esti- mated background can be subtracted from the current frame, the background image needs to be generated from the background modelling stage, which represents a very important step for the whole background subtraction algorithm. The following sub- sections describe the most practical techniques employed in background modelling, which fall into two main categories: recursive and non-recursive methods.
Figure 2.4: A simplified version of the camera-based fall detection system, depicting the image processing stage.
2.3.1
Non-recursive methods
Non-recursive methods utilise a sliding window function, which keeps a certain num- ber (N ) of frames in a buffer; based on these buffered frames, an estimation for the background model is effectuated. This technique is adaptive, as the model is deter- mined by the previous N frames and is not influenced by the frames that appeared
CHAPTER 2. IMAGE PROCESSING AND TRACKING METHODS
before these buffered ones. However, this buffer function needs a large amount of memory, especially if a large buffer is utilised [28]. The following three subsections describe the various non-recursive techniques.
2.3.1.1 Frame differencing
Frame differencing is a very simple background modelling method which considers the frame at time t-1 as the background model. This implies that the background is modelled on the previous frame. In order to obtain the foreground or the mov- ing object, the background model is subtracted from the actual image using some threshold value [29]. This method is sensitive to the threshold value chosen to dis- tinguish the foreground from the background, which is the only factor that can affect the result. Its advantages are low computational load and memory requirements, but it does not provide precise results and is sensitive to noise. This represents a major issue, especially when the person does not move for a brief period of time (more specifically for one frame), because the system will then interpret the person as part of the background.
2.3.1.2 Median filtering technique
Another effective background modelling technique is median filtering [30]. In this background subtraction approach, the background value for each pixel is determined as the median of that pixel in all the frames found in the buffer. Although this technique can give good results while employing low computational power, it does require a certain amount of memory (N x framesize) [29].
2.3.2
Linear predictive filter technique
In the third non-recursive approach, the current background model estimate is de- termined by applying a linear predictive filter to the pixels of the buffered frames.
CHAPTER 2. IMAGE PROCESSING AND TRACKING METHODS
In this model, the filter coefficients are calculated from the sample covariances for each frame time [28]. This technique is not reliable in real-time systems because of its lengthy computations.
2.3.3
Recursive methods
The main difference between non-recursive and recursive methods is that the latter do not employ a buffering system consisting of previous frames, but instead up- date the background image in a recursive manner. The advantage of using such technique is that there is only one frame which is stored and this frame or image is updated whenever a new frame is received [31]. However, when the background image becomes faulty due to some erroneous factor, this fault will take a long time to disappear from the estimated background image. Therefore, recursive methods are less adaptive than non-recursive ones. The following subsections describe the various recursive techniques.
2.3.3.1 Running average method
The running average recursive technique is a simple and fast modelling algorithm which does not require high memory. However, this method does not give accurate results and this normally depends on the type of application (where alpha can be fine-tuned so that it can be acceptable). Overall, it has simple computation and memory requirements, at the expense of poor accuracy [28].
2.3.3.2 Approximated median filtering technique
The approximated median filtering algorithm was developed by McFarlane and Schofield to track animal motion as part of a background modelling method [32]. When a pixel in the current frame has a greyscale value greater than the equivalent
CHAPTER 2. IMAGE PROCESSING AND TRACKING METHODS
by one. Conversely, when the pixel value in the current frame has a value lower than the pixel value of the background estimate, then the latter is decremented by one. Therefore, when this function is applied to a background model, it converges to an estimate where half of the input pixels are greater than the background model while the other half are less. Approximated median filtering can achieve good results with percentage accuracy close to that of algorithms requiring higher complexity. The amount of memory space required is also low and the technique is robust. Its only drawback is its slow adaptation to major changes in the real background [28]. 2.3.3.3 Kalman filtering
This background modelling technique relies on the assumption that the best in- formation on a system state is obtained through estimation [33]. The literature describes several approaches to making this type of estimation [28], most of which use the luminance, the luminous intensity together with its temporal derivatives or luminous intensity with its spatial derivatives. The advantage of this method is that the gain matrix can change from fast to slow adaptation according to whether a pixel is in the background or foreground. The disadvantage of Kalman filtering is that it leaves long trails behind the moving body or object.
2.3.3.4 Mixture of Gaussians
Finally, the mixture of Gaussians is a popular method of background modelling. Un- like the Kalman filtering method, which can track only one Gaussian, this technique can usually track three to five Gaussian distributions simultaneously [34, 35, 36, 37]. Its advantages are that it normally results in good accuracy and needs a fairly mod- est amount of memory space. A distinct difference from other methods is that the mixture of Gaussians does not utilise one image of values as its background model; rather, each pixel is modelled by a number of Gaussians, which represent the prob-
CHAPTER 2. IMAGE PROCESSING AND TRACKING METHODS
ability distribution function F. While this technique produces accurate results, its disadvantages are complexity of computation and high sensitivity to abrupt changes in illumination.