44
1.5. SYSTEM INTEGRATION AND FPGA PROTOTYPE
71
With a low threshold value where less overlapping Gaussian distributions are
regarded as the same, more savings could be achieved. However, more noise
is generated due to increasing mismatches in the matching block. Fortunately,
such noise is found non-accumulating and therefore can be reduced by later
morphological filtering [35]. Figures 1.20,1.21,1.22,1.23 show the memory
bandwidth savings over frames with various threshold values. The simula-
tion results are obtained from matlab, with four video sequences provided by
AXIS [34] are evaluated for both indoor and outdoor scenes. The sequences are
selected to reflect a range of real-world background environments with possible
difficulties for many segmentation and tracking algorithms. The scene “stairs”
comprises people moving up and down the stairs randomly. Gradual illumi-
nation changes together with shadows are the major disturbing factors. The
scene “hallway” focus on the scenarios with people moving closer or further
away from the camera. The foreground object size is varying over time. The
scene “trees” address the issue of quasi-static environments where a swaying
tree is present as the dynamic background object. The scene “parklot” presents
aenvironment with walking people, moving cars of different size. Gradual illu-
mination as well as waking foreground object are also within the focus. It can
be seen from the sequences, memory reductions scheme works robustly within
different real-world environments with variation only in the beginning due to
varied foreground actives. During initialization phase, only background pixels
are present, which exhibit high similarity within neighboring pixels. With fore-
ground objects entering the scene, part of Gaussian distributions are replaced,
which results in the decrease of number of similar Gaussian distributions. The
trends will continue until it reaches a certain point where most pixel locations
contains a foreground distribution. The decrease will flattens out in the end
since more foreground objects always replace the distribution that represent
aforeground pixel. Foreground objects activities can vary in different video
scenes, e.g. continuous activities in figure 1.20(a) where people going up and
down the stairs all the time, and the two peak activity periods around frames
600 − 900 and frames 2100 −2500 in figure 1.21(a), where people walking by in
two discrete time period. In the long run, the bandwidth savings tends to sta-
bilize (around 50%−75% depending on threshold value) after the initialization
phase. Another test sequence is also experimented in our lab. Similar results
are observed as shown in figure 1.24. The quality of segmentation results before
and after morphology are shown in figure 1.25, where it is clear that memory
reduction comes at the cost of segmentation quality. Too low threshold value
results in clustered noises that would not be filtered out by morphological fil-
tering. In this implementation, a threshold value of 0.8 is selected, combined
with wordlength reduction scheme, a memory bandwidth reduction of over 70%
is accomplished. To evaluate long term effects of memory bandwidth reduction
scheme, FPGA platform is required to collect data in real time.