• No results found

Equation 8-5. Chi-squared histogram testing

8.3 Operational Parameter Analysis Tracking Trials

8.3.2 Chi-Squared Tracking Models

Chi-Squared comparison methods employed for histogram tracking models have produced sensitive results, meaning that parameter values for the models produce wildly fluctuating precision scores. It was not possible to select only tracking models which scored high precision scores.

Figure 8-11. Test results of the HSI Histogram segmentation tracking model Success Rate (Precision)

100% 99.7% 95% 68% 50% 40% 30% 20% 10% 0%

2D Histogram

(base:30) 3 14 0 1 1 0 0 0 0 3

2D Histogram

(base:18) 12 6 0 0 0 1 0 0 0 3

Success Rate (Accuracy)

100% 99.70% 95% 68% 50% 40% 30% 20% 10% 0% 2D Histogram

(base:30) 0 0 4 3 5 4 1 0 1 4

2D Histogram

(base:18) 0 0 0 1 3 1 0 4 2 11

Analysis of the chi-squared 2D Histogram comparison results, shown in Figure 8-12, is only able to associate precision scores due to the variability of the data. It is difficult to identify a specific histogram bin size that has the best outcomes for multiple models. Peak results for some bin sizes result in the worse-case scenarios for other models. Four models peak at bin size 24, with the remaining models having precision scores in the top third of their overall variability.

The HSI Histogram model results shown in Figure 8-13 are also relatively consistent across all models. While most models display effective results towards the extreme left of the graph, the best results have occurred with a histogram bin size of 18. However, investigating the full set of results, and as the histogram bin size increases, so does the

Figure 8-12. Test results of the 2D Histogram chi-squared tracking model

run-time for the model. With a histogram bin size of 10, the tracking model is easily a real-time mode, but with a bin size of 18, all of the models exceed the time required to be considered a real-time model. For this reason, HSI Histogram tracking models using the chi-squared method employ a bin size of 10.

8.4 Summary

Tracking objects within digital video streams requires consistent and reliable detection of the objects and the means to continually recognise the object as it progresses through the video scene. Many current tracking systems work well, but only in an offline mode where the extensive processing requirements are not impacting the output. For systems such as Augmented Reality, real-time object detection and tracking is necessary or the system becomes ineffectual for the users. Isolating foreground and background regions of the incoming video frames involves segmenting the images based on rules for the model. Colour histograms are capable of recognising objects by segmenting video images according to the histogram colour frequency distribution.

Applying colour histogram segmentation to classify foreground and background objects is effective. For tracking purposes, the histogram segmented data set requires a means to select the appropriate foreground object, and ensure it is able to track the object through the video sequence, in real-time. Segmenting video sequences using the colour zone of a selected region of the object of interest and applying tracking techniques to the foreground objects appears to be effective.

As demonstrated within Chapter10.2, Segmentation Object Detection, histogram segmentation with Centre of Mass (CoM) hotspot tracking is capable of maintaining a reliable track of the object of interest; the method, however, does suffer from jitter. The tracking box can move plus or minus two pixels off the actual tracked location, causing unwanted jitter which can be disconcerting within an Augmented Reality environment. To improve jitter, histograms using the HSI colour space instead of the RGB colour space should theoretically improve these issues as a result of the psychophysical and spatial relationship between colours.

Bin size has been shown to affect various attributes of the colour histogram such as: • increasing the number of bins (reducing their size) increases the histograms

• reducing the number of bins to include in hypothesis testing has been shown to have only minor effects on accuracy [210] matching. The image information is carried by the largest bins, and as such minimal impact on accuracy is

achieved by including the tail of the histogram distribution.

The size and number of bins has some factor over histogram intersections, but for colour process, it has been shown that not all bins are required to achieve near perfect matching [148, 214]. The tracking precision for both the two-dimensional histogram model is clearly demonstrated as an improvement over the HSI histogram model. As measured previously, the new model is also more efficient with ICT resources. From the tracking results, the new two-dimensional histogram model performance is optimum for segmentation tracking when the square pyramid base size is 18, while the model is optimal for Chi-Squared tracking when the base size is 24. Due to the statistical problems with the ‘goodness-of-fit’ test, the chi-square ‘test-of-homogeneity’ test should be used to prove histogram hypotheses.

The contributions of this research have demonstrated a new colour image histogram model, which successfully creates a unique signature for object indexing or identification. The new model has been shown to provide object detection through image segmentation methods and histogram matching techniques. From the contribution, an improvement in operating speed for object detection and tracking systems has occurred, supporting the requirements of Augmented Reality within a Remote Access Laboratory environment.

Further Computer Vision tracking metrics are defined and discussed in Section 11.2, Experimentation Methodology and Section 11.6, Markerless Object Tracking.

9

9Image Object Detection Output Attributes for Signature

Matching

This chapter describes the various attributes available from CV object detection methods, and the means to compare and assess their level of match against objects of interest. A blend of current methodologies has been investigated to provide a CV object matching model which is suitable for Augmented Reality systems in a Remote Access Laboratory environment.

Computer Vision object detection has no purpose without a means to identify the object; if not as a specific named object, then as an object that is similar to an object of interest. Taken in the context of Computer Vision object tracking systems, this chapter defines the methods available for image and object attribute comparison for the purpose of object tracking, and presents a new novel approach to the object matching process. Computer Vision object detection systems provide many varying attributes which are to be associated with a chosen object of interest. For Augmented Reality processes within the Remote Access Laboratory environment, comparison of selected attributes must be performed in a manner that reduces the level of Type I and Type II errors (shown in Table 6-1), ensuring that follow-on systems, such as object tracking sub- processes, reliably locate similar attributes from frame-to-frame. Comparison processes must also function in a real-time manner. The failure of a signature matching system places a burden upon application systems such as Augmented Reality. Signature matching failures produce out-of-sync and out-of-alignment virtual images, wrenching the user out of the immersive environment, and drawing the users attention to the technology instead of the environment [63]. From the contributions within this chapter, methods are presented which impart a metric to assess object attribute collections. The

metrics allow comparison of attribute collections to ascertain the likelihood that two collections are similar.

This chapter is structure as follows: Section 9.1 summarises the requirements of Computer Vision object detection mechanisms. Object detection models and the object signature construction methods are described in Sections 9.1.1 to 9.1.7. Assessment of the object detection signatures are presented in Section 9.2, and the summarised in Section 9.3.