The research framework and organization are shown in Figure 1.5, which describes the required functions of a road-side radar system. Chapter 2 will introduce the related reviews and Chapter 3 introduces an overall system architecture including hardware and software design. After that, three main algorithms will be revealed in Chapter 4, 5, 6, respectively. The aim of Chapter 4 is to develop a background adapter to accommodate environmental changes, such as an intruding parked car or
Thread
Chapter 4 Chapter 5 Chapter 6
Update
Figure 1.5: The organization of this study
temperature variation. The goal of Chapter 5 is to propose two novel features, span and conflict information, to work with one-dimensional Gaussian Mixed Model, and a classification EM algorithm is also introduced to develop a lane boundary estimator.
A two-dimensional Gaussian Mixed Model is proposed in Chapter 6 and employed to develop the learning model based on FMCW radar data, and the model could be used for classifying small and large vehicles in multi-lane environments simultaneously.
Finally, Chapter 7 and 8 summarize numerical results in the field test and concluding remarks, respectively.
Chapter 2
Literature Review
In this section, the related literature reviews are introduced. The first part is to review the design of lane boundary estimator for multi-lane environment. However, there is few literatures associated with this topic, since this function may not have widespread applications in real-world. So far there are only two patents that are claimed by the current marketing road-side radar manufacturers.
The second part is to review the design of radar-based vehicle classifiers no matter forward or road-side installations. In addition, the feature selected for classifying vehicles are also reviewed and discussed in detail.
2.1 Literature Review for Lane Boundary Estima-tion
Essentially, developing an on-line automatic traffic lane estimator can be considered as a clustering problem, as is demonstrated in the patent [11]. This patent uses the same frequency-domain information mentioned above to decide the lane position.
After obtaining the similar frequency-domain domain information shown in Figure
13
1.2, the patent utilizes the peak information to form the vehicle feature. The bin with the maximal amount of energy is termed the peak, and the bin with most peak count is used to represent the corresponding vehicle position. In other words, a single passing vehicle contributes just one vehicle count to the corresponding bin. When the sufficient vehicle data are gathered, Gaussian mixture model (GMM) is introduced to model the probability density of lane position. The peaks of probability function describe the lane center, while the low density parts of GMM distribution represent lane boundaries. The description of lane boundary of radar is just to verify that the passed vehicle belonged to which lane, in fact, the description of lane position may not correspond to the real-world lane position.
However, representing a vehicle position by adding just one vehicle count to a bin may involve bias risk, and such an expression may easily result in misleading and incorrect lane outcomes. The examples shown in Figure 2.1 are selected from real-world data, where Figure 2.1(a) and (b) show the peak counts when small vehicles passed through the detection zone in lanes 1 and 2, respectively. According to the vehicle feature proposed in the patent [11], both Figure 2.1(a) and (b) added one vehicle unit to bin position 23, respectively. In other words, even if Figure 2.1(a) and (b) had different bin spectrums arising from the different vehicles in the different lanes, both cases provided the same feature expression that possibly caused ambiguity in separating the two lanes. The detailed outcomes of the patent [11] will also be exhibited in the latter section.
The patent [12] adopts another approach to describe the lane position based on a set of lane center variables, and utilizes displacement information to adjust the lane center. Restated, every reflected signal when a vehicle passed through the detectable
15
Small vehicle in Lane 1
0
Small vehicle in Lane 2
0
Figure 2.1: Examples illustrating the accumulated peak count scenario arose from two vehicles passed through the detection zones in two lanes. (a) indicates the accu-mulated peak count when a small vehicle passed through the detection zone in lane 1, it reveals that position 23 has the maximum bin count. (b) indicates the scenario where a small vehicle passes through the detection zone in lane 2, but position 23 still has the maximum bin count.
area could comprise zone center information. If the newly formed information is sufficiently close to the existing zones, the nearest zone is updated; otherwise, a new zone is defined based on this new formed information. However, the patent [12] easily result in misleading and incorrect lane number outcomes because of the fixed lane width, consequently, follow-up human interventions are usually required. Although this patent has the advantage of not requiring lane number input in advance, it has the weakness that incorrect estimates of lane number make it necessary to spend a long time adding/removing corresponding lanes and adjusting lane positions.
Both of the above patents transformed reflected signals arising from a passing ve-hicle into single-value information (such as bin position, lane center), as may commit bias in estimating the lane position. The main difference between the two patents is that the former is a batch algorithm, while the latter is a real-time algorithm.
This study focuses on designing an online traffic lane estimator for a road-side radar detector in multi-lane environments. New features (such as span information and conflict information) are introduced to describe lane positions such that the am-biguity in feature expression can be eliminated. Additionally, the variant of EM algorithm is proposed to achieve classification and reduce the variance of each Gaus-sian component.