Shape Memorization and Recognition of 3D Objects Using a Similarity-Based Aspect-Graph Approach

Short Paper

Shape Memorization and Recognition of 3D Objects

Using a Similarity-Based Aspect-Graph Approach

JWU-SHENG HU AND TZUNG-MIN SU

Department of Electrical and Control Engineering National Chiao Tung University

Hsinchu, 300 Taiwan

This work proposes an incremental combinational algorithm to generate the proto-type of a 3D object using 2D images randomly sampled from a viewing sphere. Similar-ity-based aspect-graph, which contains a set of aspects and prototypes for these aspects, is employed to represent the database of 3D objects. Furthermore, the proposed algo-rithm is based on low-level features and similarity measures between the features. In this work, the Fourier descriptor and point-to-point lengths are adopted as features, and three similarity measures, called the 1-norm, 2-norm, and K-L distance, are adopted to extract characteristic views. The effectiveness of the proposed algorithm is demonstrated by experiments with an updating mechanism.

Keywords: aspect-graph, fourier descriptor, object recognition, shape memorization, simi-

larity measure

1. INTRODUCTION

Object recognition is an important topic in computer vision owing to its applicabil-ity in many applications, including mobile robot localization and navigation, visual ser-voing, surveillance and military applications. Although various approaches have been developed to solve the recognition problem [1, 2], object recognition remains a difficult task in the field of computer vision, especially for 3D object recognition. Variations in viewing direction and angle, illumination changes, scene clutter and occlusion make 3D object recognition difficult and impractical in real-world applications.

3D object recognition differs from the variations of the position and type of the il-lumination source, or the relative position of observer and object. Therefore, some high- level theorems of 3D object perception have been studied to solve the above-mentioned weakness, improving the practicality of the 3D object recognition task [3]. Existing theo-rems about the high-level 3D object perception can be categorized as object-center and viewer-center representations based on the coordinate system [1], and as volume-based (or model-based) and view-based representation based on the constituent elements [4]. A viewer-centered representation describes the parts of an object relative to a coordinate Received April 20, 2007; revised August 23, 2007 & March 3, 2008; accepted April 10, 2008.

Communicated by Jenq-Neng Hwang.

8-system based on the observer. A view-based representation represents a 3D object with a set of object views.

The traditional aspect-graph method [5, 6] assumes that an object belongs to a lim-ited class of shapes, and that characteristic views can be extracted using prior knowledge of the object. Cyr and Kimia [7] presented a similarity-based aspect-graph method, using curve matching and shock matching as the similarity measures between two views, to improve the practicability of aspect-graph representation for extracting the characteristic views. However, the process of updating the aspect-graph representation is inconvenient, since the training views of an object in [7] are sampled at 5-degree increments. To im-prove the flexibility of the update mechanism, a similarity-based aspect-graph represen-tation that can be constructed with object views sampled at random intervals is studied in this work. Furthermore, the object representation becomes increasingly accurate after gathering more new object views with limited increasing search space using the proposed combinational algorithm. Additionally, this study utilizes three different similarity meas-ures from the similarity measurements used in [7], called 1-norm, 2-norm and K-L dis-tance, to reduce the computing time. Furthermore, to represent rigid objects efficiently, Fourier descriptors of the sampled points on the object contour and point-to-point lengths are calculated based on the shape contour in this work. Fig. 1 illustrates the block dia-gram of the overall scheme, where T1 denotes the minimum number of sampled 2D

views adopted to extract the object prototype.

The remainder of this paper is organized as follows. Section 2 presents the proce-dure of extracting the main (Fourier descriptors) and assistant (point-to-point lengths) features, which are utilized to measure the similarity between two views. Section 3 de-scribes the novelty of this work, namely the construction and updating of similarity- based aspect-graph representations from a set of objects views sampled at random inter-vals, and the procedure of recognizing 3D objects from 2D object views. Section 4 pre-sents experimental results to demonstrate the performance of the proposed method. Con-clusions are finally drawn in section 5.

No

An 2D view sampled at a random interval

Edge Detection and Contour Extraction

Edge Detection and Contour Extraction The number of sampled views > T1 A similarity-based aspect-graph combinational algorithm 2D test view Foreground Detection

Build the main and assistant features

Collect the characteristic views of an object

Similarity Measure Foreground

Detection

Build the main and assistant features

Building Object DateBase

Extract Test Image Features

Yes

Recognition Results

Fig. 1. Basic workflow of the proposed framework; T1 denotes the number of sampled views re-quired to build the aspect-graph representation of an object.

2. FEATURE EXTRACTION

2.1 Foreground Detection and Contour Extraction

Shadows and highlights should be removed before extracting the object features to eliminate the effects of lighting. A robust background subtraction framework from our

previous work [8, 9] is adopted to extract the foreground regions with the consideration of shadows and highlights. This study utilizes the shape feature to measure the similarity between two object views. To extract the shape information from the foreground object, Canny edge detection [10] is adopted to obtain the shape edge, and Gradient Vector Flow Snake [11] is then applied to extract the contour information. The contour information is included in a set Z, which consists of N points zi, where zi can be described as a complex

form as in Eq. (1).

Z = {zi} = {xi + jyi}, 0 ≤ i < N ( 1 )

2.2 Generation of the Main and Assistant Feature

To avoid variations in shift and scale, the points inside the set Z are re-sampling us-ing Eq. (2). The Fourier transform is applied on Z to derive the Fourier descriptors. The first T2-magnitude parts are extracted as the main feature to describe the object shape

without the variations on the high-frequency noises. The method for extracting the main feature Fm is described as in Eq. (3).

{ } {(_{i i i}} { [(_{c i c}) ( _{i c})]/ }

Z= z_ = x_ + jy_ = L x −x + j y −y L ₍₂₎

where 0 ≤ i < N, L denotes the real contour length, and Lc denotes the expected contour

length.

Fm = {| ft |, | fN−t |, 0 ≤ t ≤ T2} (3)

where | ft | and | fN−t | denotes the magnitude part of Fourier descriptor at frequency t.

Moreover, to consider take the details of an object, the lengths between each two points of the set Z are calculated as the assistant feature Fa, which is described as Eq. (4).

2 2 1 1 1 { } {|| ||} { ( ) ( ) } a i i i i i i i F = l = z − z₋ = x −x₋ + y −y₋ (4) where 0 ≤ i < N. 3. FEATURE EXTRACTION 3.1 Similarity Function

To determine the similarity between two contours, a similarity measurement has to be applied on the extracted two contours, which each consist of N points. Assume that the features extracted from two contours are respectively U = {u0, …, ui, …, uN−1} and V

= {v0, …, vi, …, vN−1}. The three similarity measures are then calculated using Eqs. (5)-(7).

• 1-norm distance: D_1−norm( , )u v =

∑

• 2-norm distance: 1 2

2 norm( , ) _iN₀ ( i i)

• K-L distance [12]: 1 0 1 1 0 1 0 0 1 1 0 1 0 0 0 1 ( ) ( ) ( || ) ( ) log( ) ( ) log( ) ( ) ( ) , , , ( ) ( ) ( ) 2 N KL t N N sum i sum i sum sum i i p t p t D p p p t p t m t m t U V p p u u v v u v p t p t m t − = − − = = ⎛ ⎞ ≈ _⎜ ⋅ + ⋅ _⎟ ⎝ ⎠ = = = = + =

∑

∑

∑

3.2 Generation of Aspects and Characteristic Views

Cyr and Kimia presented a combinational algorithm [7] (includes two criterions, called local monotonicity and object-specific distinctiveness of aspect views) to generate the aspects and characteristic views via collected 2D views sampled at 5-degree incre-ments in order. If more views of an object are captured to improve the object representa-tion in the work of [7], the total views of an object have to be resorted in order of cap-tured angles to extract the new characteristic views and aspects. The first criterion in the work of [7], called the local monotonicity, might not be matched when the object itself is symmetric in some level or the object becomes symmetric in the feature space.

This work proposes an incremental combinational algorithm motivated by the work of [7] to solve the above problems. The aspects of 3D objects can be extracted using 2D images sampled at random intervals. The object representation becomes increasingly detailed using new captured views. Besides, characteristic views are extracted without re-calculating the similarity measures by re-sorting total collected views. Such an incre-mental combinational algorithm provides a supervised learning ability for related appli-cations in an uncontrolled environment. For example, a home robot memorizes a 3D ob-ject well when it gets more and more 2D views of the 3D obob-ject at different times. Fur-thermore, the proposed algorithm extracts characteristic views by measuring the similar-ity between a new 2D view and the extracted characteristic views. The relationship be-tween the sorted neighbor views is not considered in the proposed algorithm. Thus, a local minimum region caused from the first criterion in the work of [7] is not necessary for extracting the aspects. Although the proposed approach cannot verify the test view with the specific view of an object, it improves the flexibility of building aspect-graph representation, and reduces the computing time of updating object aspects.

Assume that V_newn denotes the new sampled view of the nth _object; n m

C denotes the

mth characteristic view of the nth object; min ₁

n m

C ₋ and min ₁

n m

C ₊ denote the neighboring views with a minimum distance between n

new

V , and mmin _{denotes the index of the aspect that has} the minimum distance with n

new

V . Aspects and characteristic views are then imposed using three steps, shown as steps A-1 to A-3. Fig. 2 shows the flowchart of the proposed as-pect-graph representation.

Step A-1: When the number of existing aspects equals zero, Vnewn is considered as a

Total Aspects = 0? A new 2D view

Total Aspects = 1 or 2?

Add a new aspect and set the new view as the characteristic view of the

new aspect

Combine the new view into the aspect. Keep the characteristic view of the aspect

as the same. min m min m (8) is satisfied ? (9) or (10) is satisfied ? (11) is satisfied ? (8) is satisfied ?

Add a new aspect between aspect and

aspect and set the new view as the characteristic view of

the new aspect

min

m

1 min+

m

Add a new aspect between aspect and

aspect and set the new view as the characteristic view of

the new aspect

min m 1 min− m Yes Yes Yes No No Yes Yes No No Yes No No Step A-1 Step A-2 Step A-3

Fig. 2. The procedure of the proposed combinational algorithm procedure.

Step A-2: When the number of existing aspects equals 1 or 2, n new

V is combined into one existing aspect and the characteristic views of the aspect remains the same if Eq. (8) is satisfied. Otherwise, a new aspect is generated, and n

new

V is considered as the characteris-tic view of the new aspect.

3 all min (Cm new, m)

d V C <T (8) where T3 denotes a predefined threshold value.

Step A-3: When the number of existing aspects is greater than or equal to 3, a new aspect is built, and V_newn is regarded as the characteristic view of the new aspect if Eqs. (9) or (10) are satisfied. Otherwise, V_newn is combined into aspect mmin_{, which retains its} previ-ous characteristic view.

4 all min (Cm new, m)

d V C >T (9) min

3 _all min ( , ) 4 and ( , 1) 4

m new m new m

C

T ≤ d V C ≤ T d V C _± >T (10) where T4 denotes a predefined threshold value.

Furthermore, if a new aspect is built, then the aspect order can be determined using Eq. (11). If the similarity distance between V_newn and C_mnmin₊₁ is greater than the similarity distance between Vnewn and C_mnmin₋₁, then the new aspect is inserted between aspect mmin and aspect mmin−1. Otherwise, the new aspect is inserted between aspect mmin _{and aspect}

mmin+1_{. Therefore, the similar aspects are close to each other.}

min ₁ min ₁

( _new, _m ) ( _new, _m )

d V C ₊ >d V C ₋ (11) Terms T3 and T4 denote two predefined threshold values, where T4 > T3. The crite-rion for selecting T3 and T4 depends on the feature selection, and on the precise level for describing the object. If T3 and T4 are both small, then the criterion of combining 2D views becomes strict and thus the number of aspect becomes more numerous. Further-more, if the difference between T3 and T4 decreases, then the tolerance of difference be-tween 2D views inside an aspect decreases, thus increasing the number of aspects. Fur-thermore, T3 and T4 should be initialized manually and modified iteratively until the final number of aspect reaches an acceptable level, determined by the symmetry of the object. In this work, T31 and T41 are defined as the T3 and T4 of the main feature (Fourier de-scriptors), and T32 and T42 are defined as the T3 and T4 of the assistant feature (point- to-point lengths). Section 4 presents the values of T31, T41, T32, and T42.

3.3 Object Recognition Using 2D Characteristic Views

After constructing the aspect-graph representation of each object in the database (Fig. 2), a test view of an unknown object can be recognized using the similarity measure with the main features and the assistant features.

Step B-1: The test 2D view of an unknown object is compared with 2D characteristic views using the main features. The first T5 2D characteristic views in the database having the smallest similarity distance with the test 2D view are then preserved for further rec-ognition.

Step B-2: Suppose that T6 objects are included in the selected T5 characteristic views, the final similarity distances can be calculated from the assistant features of these T6 objects using Eq. (12). min assistant 4 3 ( ,ji mn) ( ,ji mn) ( / ) main( ) d V V =d V V + T T ×d n (12) where i j

V denotes the views of an unknown object; n m

V denotes view m of object n, and

min main( )

d n denotes the minimum similarity distance calculated using the main feature be- tween the unknown object and the nth _{object of the database.}

4. EXPERIMENTAL RESULTS

This section describes several experiments that demonstrate the effectiveness of the proposed method. A SONY EVI-D30 PTZ camera was utilized to capture object views.

Fig. 3 displays the image database contains 12 real rigid objects. The training views of each object were captured with five-degree increment intervals, and were collected as

{ ( )},

n n

d

V = V d 1 ≤ n ≤ 12, 1 ≤ d ≤ T1, with T1 views for each object. The additional views of each object were captured from the trisection-points between each five-degree point, and were collected as n { n( )},

t

V = V t 1 ≤ n ≤ 12, 1 ≤ t ≤ 216, with 216 views for each ob- ject. Furthermore, in the following experiments, T1 = 72, T2 = 25, T31 = 640, T32 = 800,

T41 = 336 and T42 = 480. Furthermore, T5 is defined as the half of total characteristic views in the aspect-graph. The computing time taken to calculate the similarity between a test view and a view in the database was about 0.005 seconds with P4 3.2G CPU and 1GB RAM.

Fig. 3. Image database comprising twelve 3D rigid objects, where object 1, object 2, …, object 12 are listed from left to right and from top to bottom.

4.1 The Efficiency of Three Similarity Measures on Rigid Object Recognition

The first experiment shows the aspect numbers extracted using the algorithm pro-posed by Cyr and Kimia [7] and the recognition results using the extracted aspects and three similarity measures, 1-norm distance, 2-norm distance, and K-L distance. The views in n

t

V were then tested with the characteristic views of each aspects using three similarity measures. Table 1 shows the results of the aspect numbers of each object. Ta-ble 2 shows the recognition results. The results in TaTa-bles 1 and 2 indicate that the as-pect-graph representation based on 1-norm distance generates the fewest aspects, and has the best recognition performance. Furthermore, the computing time of 1-norm distance was also the lowest among the three similarity measures. Therefore, for brevity, 1-norm distance was adopted to demonstrate the performance in the following experiments.

4.2 3D Rigid Object Recognition with Building Database Using 72 Object Views

In the second experiment, the efficiency of the proposed algorithm was measured using 2D views captured at random intervals. To determine the average performance of the proposed method, aspects were generated by sampling the views in 200 different ran-dom orders. For performing the efficiency of the proposed method, a comparison between

Table 1. Numbers of aspect using the main feature and aspect-graph representation in [7] with three similarity measures.

The index of the objects in the database listed in Fig. 3 Numbers of

aspect 1 2 3 4 5 6 7 8 9 10 11 12

KL distance 48 29 38 34 29 32 38 29 42 35 32 32

1-norm distance 39 35 33 42 29 32 38 23 31 37 32 40

2-norm distance 38 34 51 50 33 39 37 22 27 47 39 49

Table 2. Result of matching unknowns using 1-norm, 2-norm and K-L distance with aspect-graph representation in [7].

Recognition Rate In Top 1 Matches (%) In Top 2 Matches (%) In Top 3 Matches (%)

KL distance 97.61 98.30 98.61

1-norm distance 98.69 99.04 99.15

2-norm distance 97.34 97.69 97.80

Table 3. Numbers of aspect using the proposed method (M2) and that proposed by [7] (M1) with 1-norm distance.

Numbers of aspect

with the main feature with the assistant feature Numbers of aspect Index of objects _{M1 M2 M1 M2} object 1 39 34.7 37 38.7 object 2 35 3.8 40 14.1 object 3 33 27.8 34 14.3 object 4 42 24.8 29 22.8 object 5 29 6.9 35 11 object 6 32 9.5 41 20.1 object 7 38 2 39 8.4 object 8 23 25.6 34 31.1 object 9 31 17.1 26 25.8 object 10 37 16.2 32 17.7 object 11 32 16.6 36 23.6 object 12 40 28.8 26 19.9

the proposed combinational algorithm and that proposed by [7] is performed with the same dataset (rigid object, Fig. 3) with 1-norm distance. Table 3 shows the number of aspects, where M1 denotes the results performed with the algorithm proposed by [7] and M2 denotes the results of the proposed method. In Table 3, the number of aspects of symmetric object in M2 is fewer than those in M1, especially for objects 2, 5, 6 and 7. Moreover, Table 4 shows the recognition results. The views in V_tn were adopted as un- knowns, and tested whenever the aspect-graph representations were built each time (total 200 times). From Table 4, the average recognition results of the proposed method and that proposed by [7] are almost the same. Table 5 shows the statistical information about the mean and the standard deviation of the aspect number. The small standard deviation in Table 5 demonstrates that the proposed combinational algorithm generated the aspects without the effects of the training views order. Furthermore, Table 6 shows the statistical

Table 4. Result of matching unknowns using the proposed method (M2) and that pro-posed by [7] (M1) with 1-norm distance via main and assistant features.

Top 1 Matches (%) Top 2 Matches (%) Top 3 Matches (%) Index of objects M1 M2 M1 M2 M1 M2 object 1 95.83 98.25 97.22 99.21 97.69 99.61 object 2 100 99.97 100 100 100 100 object 3 99.07 97.71 99.07 98.96 99.07 99.39 object 4 97.22 97.39 98.61 98.73 99.07 99.34 object 5 100 100 100 100 100 100 object 6 100 99.81 100 99.96 100 99.98 object 7 100 99.78 100 99.87 100 99.89 object 8 98.15 99.35 98.61 99.67 98.61 99.78 object 9 100 99.90 100 99.97 100 99.99 object 10 98.61 97.97 99.54 98.68 99.54 98.98 object 11 100 98.44 100 99.47 100 99.77 object 12 95.37 96.83 95.37 98.17 95.83 98.65 Average 98.69 98.78 99.04 99.39 99.15 99.62

Table 5. Number of aspects using main and assistant features and the proposed aspect- graph generation with 1-norm similarity measure on rigid object recognition, where “Std.” denotes standard deviation.

The index of the objects in the database listed in Fig. 3 Numbers of aspect _{1 2 3 4 5 6 7 8 9 10 11 12} Mean 34.7 3.84 27.8 24.8 6.9 9.5 2.0 25.6 17.1 16.2 16.6 28.8 main feature Std. 1.82 0.56 1.62 1.07 0.82 0.96 0.49 1.28 1.79 1.24 1.00 1.04 Mean 38.7 14.1 14.3 22.8 11.0 20.1 8.4 31.1 25.8 17.7 23.6 19.9 assistant feature Std. 2.51 1.67 1.81 1.79 1.17 1.94 1.09 2.07 2.18 1.68 1.81 1.59

Table 6. Recognition rates using main and assistant features and the proposed aspect- graph generation with 1-norm similarity measure on rigid object recognition.

The index of the objects in the database listed in Fig. 3

Recognition Rates 1 2 3 4 5 6 7 8 9 10 11 12 Avg. Mean 98.3 100 97.7 97.4 100 99.8 99.8 99.4 99.9 98.0 98.4 96.8 98.8 Top 1 (%) Std. 0.96 0.18 1.35 0.98 0.00 0.00 0.24 0.64 0.23 0.92 0.83 0.70 0.59 Mean 99.2 100 99.0 98.7 100 100 99.9 99.7 100 98.7 99.5 98.2 99.4 Top 2 (%) Std. 0.64 0.00 0.90 0.67 0.00 0.13 0.21 0.42 0.11 0.70 0.49 0.65 0.41 Mean 99.6 100 99.4 99.3 100 100 99.9 99.8 100 99.0 99.8 98.6 99.6 Top 3 (%) Std. 0.44 0.00 0.62 0.49 0.00 0.09 0.20 0.33 0.07 0.59 0.31 0.50 0.30

information about the recognition results, including the mean and the standard deviation of the recognition rate. From Table 6, the proposed aspect-graph generation is efficient owing to its high recognition rate and low standard deviation.

Table 7. The result of the numbers of aspect using the main and assistant features and the proposed aspect-graph representation in with 1-norm similarity measure.

The index of object in the database

Numbers of aspect 1 2 3 4 5 6 7 8 9 10 11 12 D18 14.11 3.40 11.98 10.13 5.32 6.36 1.58 12.64 8.80 8.43 8.73 11.10 D36 22.86 3.60 18.83 16.15 6.23 8.01 1.80 19.16 12.53 12.17 12.48 18.29 D54 29.52 3.74 23.95 20.96 6.645 8.94 1.92 23.24 15.20 14.53 15.06 24.05 D72 34.66 3.84 27.83 24.75 6.87 9.47 2.04 25.62 17.14 16.16 16.62 28.75 D90 39.28 3.99 28.68 25.99 7.14 9.83 2.14 27.32 18.04 17.37 17.86 30.99 D108 43.28 4.07 29.50 27.14 7.36 10.12 2.26 28.67 18.90 18.50 19.06 33.12

4.3 3D Object Recognition with Updating Database Using 18, 36, 54, 72, 90, 108 Views

The proposed method can construct the aspect-graph representation with sampled views at random intervals, making an updating mechanism that integrates the database using new collected views practicable. In this experiment, 18 random views sampled from

n d

V were first utilized to construct the coarse aspect-graph representation of each object, called D18. Eighteen additional random views were then taken from the remaining views in V to update the coarse database _dn D18 to a more accurate database, called D36. Simi-larly, D54 and D72 were constructed using views in the remainders of V_dn. Moreover, D90 and D108 were further constructed with extra random views sampled from V_tn. Table 7 shows the average aspect numbers of each rigid object from 200 runs. Although the as-pect numbers increased when new views were employed to update the coarse database, the number of stored views was still much smaller than the number of original views. Fig. 4 shows the results of recognition rate using coarse to fine databases, and Fig. 5 shows the results of the standard deviation of recognition rate. As shown in Fig. 4, the recogni-tion rate rises when the aspect-graph representarecogni-tions are trained from more object views. Moreover, the stability becomes better according to the decreasing standard deviation. Therefore, the proposed method is demonstrated to be useful for updating the aspect- graph representations without re-sorting the overall collected views, or re-calculating the overall similarity measures.

5. CONCLUSIONS

This study proposes a flexible combinational algorithm to build the aspect-graph representation with 2D images sampled at random intervals for recognizing 3D objects. Fourier descriptor and the length between each point on the object contour are adopted as the main and assistant features, respectively, to measure the similarity between each ob-ject view using three similarity measures, namely the 1-norm, 2-norm and K-L distances. Although the relationships between the aspects are lost, the number of aspects for the symmetrical objects is reduced and flexibility of the database update mechanism in-creases without the need to reorder the overall collected views. Therefore, the proposed algorithm improves the computing time while updating the database. A new view of an object can be adopted to increase the integrity of the database using the proposed algo-rithm. Furthermore, to make the databases easy to construct, foreground detection with

shadow and highlight removal is adopted to extract the interested object. Although the combination of proposed main feature and assistant feature is valuable for representing objects, the proposed method has a high computation requirement, and improving its efficiency is a topic for future work.

Fig. 4. Recognition rates of coarse and fine databases (D18, D36, D54, D72, D90 and D108), calculated with 200 results.

Fig. 5. Standard deviations of recognition rates using coarse to fine databases (D18, D36, D54, D72, D90 and D108), calculated with 200 results.

REFERENCES

1. G. Peters, “Theories of three-dimensional object perception-a survey,” Recent

Re-search Developments in Pattern Recognition, Transworld Research Network, Kerala,

India, Vol. 1, 2000, pp. 179-197.

2. G. Mamic and M. Bennamoun, “Representation and recognition of 3D free-form ob-jects,” Digital Signal Processing, Vol. 12, 2002, pp. 47-76.

3. V. Blanz, M. J. Tarr, and H. H. Büulthoff, “What object attributes determine canoni-cal views?” Perception, Vol. 28, 1999, pp. 575-600.

4. I. Weiss and M. Ray, “Model-based recognition of 3D objects from single images,”

IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 23, 2001, pp.

116-128.

5. J. J. Koenderink and A. J. van Doorn, “The singularities of the visual mapping,”

Bio-logical Cybernetics, Vol. 24, 1976, pp. 51-59.

6. I. Shimshoni and J. Ponce, “Finite-resolution aspect graphs of polyhedral objects,”

IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 19, 1997, pp.

315-327.

7. C. M. Cyr and B. B. Kimia, “A similarity-based aspect-graph approach to 3D object recognition,” International Journal of Computer Vision, Vol. 57, 2004, pp. 5-22. 8. J. S. Hu, T. M. Su, and S. C. Jen, “Robust background subtraction with shadow

re-moval for indoor environment surveillance,” in Proceedings of IEEE/RSJ

Interna-tional Conference on Intelligent Robots and Systems, 2006, pp. 4545-4550.

9. J. S. Hu, T. M. Su, and C. C. Lin, “Shape memorization and recognition of 3D ob-jects using a similarity-based aspect-graph approach,” in Proceedings of the IEEE

International Conference on Systems, Man, and Cybernetics, 2006, pp. 4920-4925.

10. J. Canny, “A computational approach to edge detection,” IEEE Transactions on

Pat-tern Analysis and Machine Intelligence, Vol. 8, 1986, pp. 679-698.

11. C. Xu and J. L. Prince, “Gradient vector flow: A new external force for snakes,” in

Proceedings of theIEEE Conference on Computer Vision and Pattern Recognition,

1997, pp. 66-71.

12. Y. Rubner, C. Tomasi, and L. J. Guibas, “The earth mover’s distance as a metric for image retrieval,” International Journal of Computer Vision, Vol. 40, 2000, pp. 99-121.

Jwu-Sheng Hu (胡竹生) received the B.S. degree from the Department of Me-chanical Engineering, National Taiwan University, Taiwan, in 1984, and the M.S. and Ph.D. degrees from the Department of Mechanical Engineering, University of California at Berkeley, in 1988 and 1990, respectively. He is currently a Professor in the Depart-ment of Electrical and Control Engineering, National Chiao Tung University, Taiwan, R.O.C. His current research interests include microphone array signal processing, active noise control, intelligent mobile robots, embedded systems and applications.

Tzung-Min Su (蘇宗敏) was born in 1978. He received the B.S. degree in Electri-cal and Control Engineering from National Chiao Tung University, Taiwan, R.O.C. in 2000. He is currently a Ph.D. candidate in Department of Electrical and Control Engi-neering at National Chiao Tung University, Taiwan, R.O.C. He is the championship of the national competition held by Ministry of Education Advisor Office in 2001. His re-search interests include background subtraction, 3D object recognition, and home-care surveillance.