Professional Documents
Culture Documents
Implementation_of_ant_clustering_algorithm_for_IR_imagery_in_wind_turbine_applications
Implementation_of_ant_clustering_algorithm_for_IR_imagery_in_wind_turbine_applications
Implementation_of_ant_clustering_algorithm_for_IR_imagery_in_wind_turbine_applications
Abstract—Interaction of avian with turbines has become an different groups, based on target features. This is an
important public policy issue, so identification and unsupervised learning method which is used for exploration
quantification of avian at turbine sites is crucial. of inter-relationships among a pool of data. In clustering
In this work, first the data is collected in terms of videos there is no need to have any priori labeling of the data. This is
recorded by an IR camera in the vicinity of wind turbine; the in contrast to classification problem which is considered as
videos are applied for detection and tracking algorithm. supervised method requiring pre-defined databases.
Features are then extracted for each detected target. Ant based
clustering algorithm (ACA) based on Lumer & Faieta with its The novelty of this work is quantification of avian
three different variations including Standard ACA, Different (bird/bat/insect) using IR images in terms of groups created,
Speed ACA and Short Memory ACA is implemented over which can then be correlated to the results obtained from
extracted features and are compared in terms of different other monitoring tools (e.g., acoustics, radar). In this work,
groups created for detected avian data. first videos recorded by IR camera are processed by target
The novelty of this work is quantification of avian detection algorithm which is based on Background
(bird/bat/insect) using IR images in terms of clusters. Also this Subtraction and Morphological techniques. Then detected
work uses Ant Clustering Algorithm which is an effective novel targets are given to a tracking algorithm which is based on
method for this application. The results of this experiment can component connectivity. Target features are extracted and
be combined with other type of monitoring systems for better
applied to the Ant Clustering Algorithm (ACA)[7]. This
understanding of avian behavior. It will also be helpful for
algorithm is a bio-inspired technique inspired from the
biologists for development of appropriate mitigation
techniques.
behavior of ants. ACA by Lumer & Faieta (LF) [8] has been
effectively used in different applications for grouping of the
data. This model can also be applied for imagery avian
I. INTRODUCTION
clustering. An important feature of this algorithm is that no
Due to numerous reports of bird/bat interaction with priori assumption on the number and size of clusters is
turbine blades and their mortality [1][2], it is important to needed, in contrast to other clustering problems such as k-
record the behavior of avian near off/on shore wind sites. means. Three different versions of LF-ACA are
Monitoring the bird/bat activity at wind-energy facilities can implemented for clustering including Standard ACA (S-
be through different techniques such as acoustic methods ACA), Different Speed ACA (DS-ACA) and Short-term
[3][4] and visual techniques (e.g. radar and IR camera) [5][6]. Memory ACA (SM-ACA). Block diagram of Fig. 1 shows
These monitoring and quantification of bird/bat activity in the various steps involve in detection, tracking and clustering
vicinity of wind turbines can be very useful for development process.
of mitigation techniques and reduction of avian mortalities.
Acoustics may be useful for species level identification
based on a database of pre-known calls of desired species. Background Feature
However the quantification of birds and bats using acoustic Subtraction Extraction
methods may not be very reliable as some birds may not call
and some bats may be hovering/calling number of times in a Image Sequence Thresholding
certain area. IR camera can be a valuable tool for
supplementing the quantification process. Videos can be
recorded over the specific period of time and they can be Noise Suppression
processed using video processing techniques to determine the
number of tracks. However separation of the counts in terms Clustering
of bats, birds or insects can be very challenging. As pre-
define images of all avian tracks are not available for training Tracking
for use in classification. The best method can be to group SACA DSACA SMACA
bird/bats via available clustering algorithms.
Clustering algorithm can arrange the set of data into Figure 1. Block Diagram of the Avian Detection and Tracking system
This work is partially supported from DOE Contract #DE-FG36-06G086096
Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on May 14,2022 at 05:27:21 UTC from IEEE Xplore. Restrictions apply.
The rest of this paper is structured as follows: First data Tracking Algorithm
acquisition is explained in Section II. Then target (1) Convert videos to frames
detection/tracking and feature extraction is described in (2) Detect blobs
Section III. Clustering including three different variations of (3) Create morphological blobs
(4) Label the blobs
LF model is presented in Section IV. Finally results and (5) Check Component Connectivity
conclusions are given. a) In case of connectivity , use same labeling
b) Else use new labeling
II. DATA ACQUISITION (6) Check for the last frame containing blob with same label
a) In case of last frame goto step 7
The camera used in this study is a thermal IR camera b) Else check new frame
(FLIR SR-19) with resolution of 704 × 480 pixels. The (7) Plot target trajectory
camera is static and it is oriented +19.75° from North, while Figure 2. Detection and Tracking process
the left side of the screen is the East and the right side is the ∑
West. The rate of the recording is 30 frames/sec (fps). 4
1 0.033
Video feeds from IR camera are stored onto a hard drive for
further data processing. The data was collected between Where is the velocity of kth bird, is the distance which
May - July 2011, starting at one hour after the sunset to one is passed by bird in frame i . The value of 0.033 is the time
hour before sunrise; as we are concerned with nocturnal of each frame including the bird, considering the camera
bird/bat migration. Because of hardware limitations, frame rate of 30 frames/second.
recorded videos are divided into small videos of fifteen
minutes each. They are then divided into number of frames. Heat: It is the average intensity value of the bird and can be
defined as:
III. DETECTION/TRACKING/FEATURE EXTRACTION 1
5
Recorded frames are used for detection, tracking, feature
extraction and clustering. Detection was implemented based Where is heat of kth bird and is the heat of bird in
on background subtraction using Running Average and frame i.
Morphological techniques [9] consisting of opening followed
by closure as[9]: Straightness Index: It is a measure of the discrepancy
between the actual path which is travelled by the bird and a
1 1 perfect straight segment and is defined as:
2 6
∑
Where is background, is foreground, α is adaptive rate
(0.05), K is binary image and S is element structure. Where is the straightness index of kth bird, is the
, , , are erosion, dilation, opening and closure operators, straight distance of bird k and is distance from starting
respectively. point to the present location of bird. The value of
straightness index is in the range from 0 to 1. Fig. 3 shows
Tracking algorithm was developed based on labeling and results from detection/tracking algorithm where image (a)
component connectivity. Eight-connected neighborhood show one track and image (b) show two tracks respectively.
was used for connectivity. Number of frames including the
bird with same label are added together to generate the
trajectory of the bird. The process of detection and tracking IV. ANT CLUSTERING ALGORITHM (ACA)
algorithm is shown in Fig. 2. Ant Colony-based clustering is a heuristic method for
clustering data inspired by behavior of ants, particularly for
Features were extracted from already tracked bird/bats. clustering of unlabeled data sets. In this method, data is
Features include size, velocity, heat and straightness index. projected from multi-dimensional space to a bidirectional
These features are very effective in separating various grid, in which ants are randomly move and pick up and
species. These features are described as follows: deposit items to and from specific cells.
Size: it is the average size of the bird, as
1
3
869
Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on May 14,2022 at 05:27:21 UTC from IEEE Xplore. Restrictions apply.
The common feature of both real and artificial ants is • Projection from attribute space to item grid
collection and drop-off. However, real ants collect corpses to • Calculate Perceived Function
make their nest cleaner or take larvae to arrange them by size. • Pick up and drop off items based on probability
While artificial ants move data items that are laid out in an
artificial environment, and specially organize them in a
B. Different Speed ACA (DS-ACA)
sorted fashion. The important feature of ACA is no priori
assumption on the number and size of clusters is needed. Similar to S-ACA ants move randomly in the item space,
however the ants have different speeds in contrast to S-ACA
Three different variations of ACA of LF model [8] are
where there was same speed for all ants. The speed is
implemented for application of avian clustering. In this distributed uniformly in the interval 1, where
application the features extracted from track of birds/bats, is the maximum speed and is defined as 6. This
including size, velocity, heat and straightness index are speed affects the probability of pick up and drops off items
applied as the attributes for clustering problem. by ants according to Eqn. 10:
Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on May 14,2022 at 05:27:21 UTC from IEEE Xplore. Restrictions apply.
Figure 4. IR camera used in this Figure 5.Location of camera
work deployment (Ottawa NWR)
(a) (b)
Data Target Tracking
Acquisition Detection
Clustering Feature
Extraction
Figure 6. Overall process
871
Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on May 14,2022 at 05:27:21 UTC from IEEE Xplore. Restrictions apply.