Comparison of Computational Speed

A comparison of the experimental computational speed and performance comparisons with those of two other studies is shown in Table 5.4. The term, “LPD,” indicates that the system only dealt with plate detection, “LPS” presents that the system only dealt with plate character segmentation and character recognition, and “All” denotes the system including plate detection, plate character segmentation, and character recognition.

Moreover, our system also outperforms commercial systems with respect to computational speeds. The VECON-VIS system [136] processes each image in less than 0.5 seconds while the SeeCar library [137] could response at 25msec per image but requiring at least 80 pixels

Table 5.3: Detection results of our plate detection module

Dim. Max./Min. Plate/Non-plate True Positives/ TPR (%) FPR(%) P late_Height Candidate # False Positives

1 320× 240 40/15 3, 756/ 341, 240 3, 718/ 87, 268 98.997 0.256 2 640× 240 60/15 4, 153/ 418, 438 4, 145/ 89, 998 99.807 0.215 3 640× 480 120/30 11, 896/1, 137, 348 11, 879/223, 636 99.857 0.197

height per typical plate. When we applied our system to the dataset, thirty-two local vehicle images used in the SeeCar library [137], all plates in the images are recognized correctly. Other systems, such as Zamir’s LPR engine [138] and CARMENR Parking License Plate Recognition Engine [139], would perform less than 20 frames per second.

Above commercial systems all proclaim that the accuracy rates are over 95%. However, the details of settings and testing environments are unavailable for detail comparisons.

(a)

(b) (c)

Figure 56: Examples of a variety of plates on a wet day: (a) input images, (b) recognized results shown in color images, and (c) enlarged recognized plates.

(a)

(b) (c)

Figure 57: Examples of a variety of plates on a cloudy day: (a) input images, (b) recog-nized results shown in color images, and (c) enlarged recogrecog-nized plates.

(a)

(b) (c)

Figure 58: Examples of a variety of plates on a sunny day: (a) input images, (b) recognized results shown in color images, and (c) enlarged recognized plates.

(a)

(b) (c)

Figure 59: Examples of a variety of plates with the pavement background of texture patterns: (a) input images, (b) recognized results shown in color images, and (c) enlarged recognized plates.

(a)

(b) (c)

Figure 60: Examples of a variety of plates captured from a device installed in a car: (a) input images, (b) recognized results shown in color images, and (c) enlarged recognized plates.

Table 5.4: Comparison results of the computational speed and the accuracy rate for plate recognition

Speed (fps) Dim. # Accuracy(%) Env.

Chang, et al., ITS’04 ˜2.5(LPD) 640× 480 1061 ˜93.7 P4 1.6 GHz

Amit, et al., PAMI’04 ˜0.3(LPS) None 520 ˜99 P3 1 GHz

Our system ˜40(All) 320× 240 22, 756 96.638 P4 3.0 GHz Our system ˜36(All) 640× 240 26, 153 97.695 P4 3.0 GHz Our system ˜38(All) 640× 480 63, 896 98.251 P4 3.0 GHz

Figure 61: Comparison of feature numbers between classifiers using only upright and both upright and skewed Haar-like features.

Chapter 6 Rejection Mechanisms for Video Sequences

In surveillance applications, search space reduction (SSR) is an essential element to effi-cient algorithms. In this study, spatial and temporal SSRs are integrated for license plate detection in video sequences; the plates could be extracted robustly and extremely fast.

Our method started from spatial SSR by a bi-level one-pass plate extraction (BOPE) algorithm developed to extract plates accurately even in complicated situations. The search space was reduced by applying information from both spatial characteristics and temporal representation in consecutive frames. We proposed to exclude repeated patterns with the similar appearances in the same location of consecutive frames, which usually include stopped vehicles or regular backgrounds. For efficiency, repeated patterns were detected only on the results of BOPE, named spatiotemporal SSR, based on a block-based mechanism by estimating the tangent distance, which is invariant to the variations in po-sitions, sizes, rotations, or brightness. To reduce the computational load, the repeated patterns and their measured invariant features could be retained for next estimation. In our experiments, the search space could be reduced up to 87.9% by the spatiotemporal SSR; all plates with the heights ranging from 24 to 40 are extracted correctly. The im-plementation of the system on a 3G Hz PC run on average 76 frames per second. The high performance suggests that the real-time goal for surveillance applications could be accomplished.

6.1 Introduction

This study develops a bi-level one-pass plate extraction (BOPE) to detect plate candidates robustly and accurately even in complicated situations, where vehicles and motorcycles may appear simultaneously in parking spaces. In these environments, plates might be close or connected with backgrounds in the captured frame. Moreover, to save computational load, we propose to reject repeated patterns with similar appearances in the same location of consecutive frames instead of detecting moving objects directly. The patterns repeated in consecutive frames would be identified utilizing the similarity measurement between the extracted plate candidates of the previous frame and those of the current frame. In LPR system, usually stand for unnecessary ones, such as stopped vehicles or regular background patterns. By the procedure of repeated region detection, redundant computations could be avoided. However, it is time consuming to compare the candidates in the current frame to all of the candidates in the preceding frame. In Section 6.3, a block-based procedure is presented to verify repeated patterns.

To detect plate candidates rapidly and effectively, the mechanism of our spatiotem-poral SSR is depicted in Fig. 62. Three major modules of the mechanism are summarized as follows:

• Plate candidates generation: This module generates a plate candidate map, Gp, by thresholding the vertical gradients of the original input.

• Bi-level one-pass plate extraction: This module extracts candidates of compact plate regions, Cand_p, by the BOPE algorithm (Section 6.2).

• Repeated region detection: This module detects repeated regions by the similarity measurement (Section 6.3) between Candp at the current frame and ones at the preceding frame.

Figure 62: Overview of the rejection mechanisms for video sequences.