Histogram Matching and Intensity Complexity

Although the edge complexity can perform well in the daytime, its capability is decreasing as time approaches evening. The main reason is that vehicles will turn on their front lamps when the weather is going darker and we only use the bottom-half part of candidate region (CR), where the lamps locate, to compute edge complexity. Because of the light comes from the lamps, the edge complexity of vehicles drop off dramatically causing its discriminative ability being not suitable for evening case. In addition, the dark lighting condition makes the color, contour and edge of vehicle become ambiguous. On the other hand, the lighting lamps become the most obvious feature of vehicle in the evening. Consequently, we use the lighting lamps as the main feature instead and utilize them to filter out false alarms.

Figure 3-13 is the example of edge image in the daytime and evening respectively.

There are two steps to identify the false alarms as illustrated in Figure 3-10. The system uses histogram matching to filter out the false alarms, mostly are ground images, and use intensity complexity to confirm the CRs being vehicle images or not. These two methods are constructed into cascaded structure inspired by AdaBoost algorithm. By cascaded structure, each step, namely histogram matching and intensity complexity, can deal with relative simple problem and filter out false alarms efficiently.

First, Figure 3-14 demonstrates two images, namely a false alarm image (ground image here) and a vehicle image, and their histogram image respectively.

(a) (b)

(c) (d)

Figure 3-14 (a) vehicle image (b) ground image (c) histogram of (a) (d) histogram of (b)

As you can see in Figure 3-14, because the lighting lamps, the histogram of vehicle is

different from ground’s. Unfortunately, the light come from lamps is also projecting on the road causing the histogram of ground can sometimes be similar with the histogram of vehicle.

Besides, the intensity of vehicle body can also be confusing because other vehicle’s lamps might project on the other vehicle. In the other words, we cannot be sure that the histogram that we computed can represent the lamps. Therefore, we do not use the entire image to compute the histogram and the range of histogram is not typical 0~255.

What we do here is similar to the idea of edge complexity. Because the lamps are usually locating at the bottom-half part of candidate region (CR) and they are approximately symmetric to the middle vertical line of CR, we utilize these properties to select sub-region to computer histogram. Figure 3-15 illustrates the sub-region we chose.

Figure 3-15 Sub-region to compute histogram

It is important that the histogram we computed is almost generated from the intensity of lamps. As Figure 3-15, we first look at bottom-half part of CR and reduce the range of left, right and button boundary. The purpose of reducing the range of sub-region is to decrease the impact of background image because the CRs always contain some background image. By reducing the range of sub-region, we can more precisely locate at the position of lamps.

sub-region to zero. This is because the lamps are approximately symmetric to the middle vertical line of CR. Therefore, we assume the pixels around the middle vertical line and within the sub-region cannot represent the lamps and we do not want to include these pixels.

Thirdly, we only include the pixel whose intensity is high than 110 when computing the histogram. By limiting the range of intensity, the difference between the false alarms and vehicle objects can be enlarged. Figure 3-16 illustrates the result of our method.

(a) (b)

(c) (d)

Figure 3-16 (a) Vehicle image (b) Ground image (c) histogram of (a) (d) histogram of (b)

After computing the histogram of each CR, we compare histograms with our two target histograms, which are selected from our training samples and tested by Equation 3-5.

(3-5)

where ρ is the correlation coefficient, X and Y are two random variables, E is the expect value operator, Cov means covariance and σ is standard deviations. The whole procedure of histogram matching is illustrated in Figure 3-17.

Figure 3-17 Procedure of histogram matching

As Figure 3-17, the structure of histogram matching is also cascaded structure. It is unrealistic to expect one histogram can handle most false alarms. Therefore we use cascaded structure to filter out false alarms as many as possible. After all, we still has next step, namely intensity complexity, and stabilizer to erase false alarms.

The next step is intensity complexity which is inspired by edge complexity and further utilizes the histogram information. In the histogram matching step, we already have the histogram of each CR and also the amount of pixels whose intensity range from 110 to 255.

confirms the CR is vehicle objects. The intensity complexity is computed by Equation 3-6.

(3-6)

where is the amount of pixels whose intensity is between 155 and 255 and n is the

number of pixels of sub-region which used to compute the histogram in histogram matching step.

The purpose of intensity complexity is to examine the existence of lamps and further confirm the CR is vehicle object. The falsification is by Equation 3-7. The idea here is similar with edge complexity. In edge complexity step, we use the complexity to represent the head of vehicles. As for intensity complexity, we use the complexity to represent the lamps. As mentioned in histogram matching, we utilize not only the intensity of bright pixels come from lighting lamps but also the position of lamps. Therefore, we only select pixels whose intensity range from 155 to 255 and location is within the assumed area to compute intensity complexity.

(3-7)

where η is empirical threshold.

The system uses the light lamps, which are the most obvious feature of vehicles in the evening, to filter out the false alarms and remain the vehicles object instead of using edge complexity. This scheme let the system can operate in evening.