Radar Target Detection Using Target Features and

Artificial Intelligence

Hai Deng and Zhe Geng Braham Himed

Department of Electrical and Computer Engineering RF Technology Branch
Florida International University Air Force Research Lab
Miami, Florida 33174, USA WPAFB, Ohio 45433, USA

Abstract— A new type of radar target detection approach In this work, we propose a new type of radar target
based on target features and artificial intelligence techniques is detection method based on distinguishing target and
proposed and investigated in this work. Traditional radar target interference features without removing interferences from the
detection in clutter and interference is achieved by removing radar receiver. Typically radar targets possess isolated-point
clutter and interference through filtering prior to target features in space domain; and no-zero single or congregated
detection. The novel approach, keeping both targets and frequencies in Doppler domain. Clutter are normally extended
interferences, recognizes and detects targets using artificial in space domain, close to zero (for ground radar) or extended
intelligence techniques based on distinguishable target and (for airborne radar) in Doppler domain. For jammers, normally
interference features. The proposed approach is equally effective
they are extended in Doppler domain and isolated in space
and more robust to environmental changes and could replace the
traditional filtering-based detection methods for all radar
domain. With these distinguishing features, targets will be
platforms. automatically detected from interferences using artificial
intelligence techniques. With recent advances in computing,
Keywords—target detection; clutter and interference; artificial VLSI, and machine learning, artificial intelligence techniques
intelligence; target features. are increasingly mature and feasible for radar target detection.
Numerous artificial intelligence techniques such as decision
trees, feature vectors and machine-learning classifiers are used
I. INTRODUCTION
for smart target detection in this work.
Radar target detection is a critical part of radar functions.
Targets can be reliably detected if the signal-to-noise ratio The rest of the paper is arranged as follow. Radar target and
(SNR) is sufficiently large [1]. Numerous techniques such as interference features will be discussed for various radar
increasing transmitting power, pulse compression and coherent platforms in section II; several artificial intelligence techniques
processing can be used to enhance target signals and thus to and schemes will be introduced for smart radar target detection
improve detection SNR. However, if radar echoes contain in section III; and finally some preliminary results are
interferences such as clutter and jammers as well as thermal presented in section IV.
noise, the interferences are considered to be a part of “color”
noise and must be removed or significantly suppressed through II. TARGET AND INTERFERENCE FEATURES FOR VARIOUS
filtering to have a large enough signal-to-interference plus RADAR APPLICATIONS
noise ratio (SINR) for reliable target detection. Therefore,
The extraction of discriminating features is crucial for
traditionally optimum target detection can be achieved by
target detection using artificial intelligence algorithms.
filtering out interferences such as clutter and jammers using the
optimum filter. Typically, MTI radars use a filter to remove Generally, some kinds of transforms are needed to generate
any stationary clutter with zero or near-zero Doppler favorable discriminating features under certain application
frequencies. Airborne Pulse Doppler (PD) radars employ a scenarios. We will consider selecting target and interference
stopband filter to remove ground clutter from specific Doppler features and their extraction methods in several typical radar
frequency bands. Airborne radars equipped with a Space-Time applications.
Adaptive Processing (STAP) system eliminate ground clutter A. Surface-Based MTI/MTD Radars
by using a space-time adaptive filter. However, all current
filtering-based target detection schemes require accurate For surface-based MTI/MTD radar, the ground clutter is
knowledge of the interferences such as their power spectra or stationary with zero-Doppler frequency. The Doppler
covariance matrices for effective interference removal; frequencies of the weather clutter might be nonzero, but are
incomplete or slightly inaccurate interference information may relatively small. The clutter is extended in space domain
lead to a significant processing loss [2]. (continuous in adjacent range bins). The target is pointed in
space domain (occupying one or two range bins); and pointed
in Doppler domain with a value that is much higher than
This work was supported by a subcontract with Defense Engineering
Corporation (DEC) for research sponsored by the AFRL under Contract
FA8650-12-D-1376 and in part by the NSF under Award AST-1443909.
clutter Doppler frequencies. The clutter and target features in Therefore, target can be detected based on its distinguishing
space and Doppler domains for ground-based MTI/MTD features without adaptive filtering.
radars are shown in Fig. 1.
Ground Clutters

Target Target

Doppler frequency
Doppler frequency

0
Range Jammer

Ground/Weather
Clutters 0
Angle
Fig. 1. Clutter and target features for MTI/MTD radars Fig. 3. Clutter, jammer and target features in angle-Doppler domain
for airborne adaptive radars
Target and clutter are separate in range-Doppler domain
and have different shape-features. The targets are detectable
based on their unique features.
D. Other Clutter and Deceptive Jammers
Discrete ground clutters generated by isolated scatters such
B. Airborne Pulse Doppler (PD) Radars as ground medal towers are pointed in feature domains, but
For traditional airborne PD radars, high PRF waveforms they are recognizable as clutter since they are stationary. In
are used to separate various kinds of ground clutter and targets addition, deceptive jammers are normally difficult to recognize
for clutter suppression filtering processing and target detection. by regular radars, but can be identified easily in feature spaces
The clutter and target features for PD radars are shown in Fig. using artificial intelligence techniques.
2.
Closing Target
III. FEATURE-BASED TARGET DETECTION SCHEMES USING
ATIFICAL INTELLIGENCE TECHNIQUES
Doppler frequency

Mainlobe
Ground Clutters Since radar echoes contain interferences such as clutter and
jammers as well as targets and thermal noise, target signals
0 should be enhanced through integration processing to ensure a
Range
large enough SNR prior to any feature-based detection for
Sidelobe reliable target detection regardless of various interferences. In
Ground Clutters Chasing Target addition, radar echo samples should be transformed to a multi-
dimensional signal domain (range/angle/Doppler) for a desired
feature recognition result. Thermal noises are still independent
Fig.2. Clutter and target features for airborne PD radars in any transformed domains. Normally SNR is enhanced
through the same data transform; and thermal noise is removed
The ground clutter and targets in Fig.2 are well separated in by hard-thresholding processing of the transformed radar
range-Doppler domains. Because clutter is extended in either echoes prior to feature-recognition processing for target
Doppler or range domain, they can be identified and detection. The general feature-based detection scheme for radar
differentiated from targets based on their different features. target is shown in Fig. 4.
Data Transform Thresholding
C. Airborne Adpative Radars Radar and Target for Noise
Feature Target
Echoes Extration Detection
Enhancement Removal
Airborne adaptive radars such as space-time adaptive
processing (STAP) radars allow low PRF waveforms to be Fig. 4. Feature-based target detection
used for adaptive suppression of ground clutter [3]. Therefore,
the covariance matrix of the ground clutter must be accurately The keys to the feature-based target detection schemes are
estimated for adaptive processing. Alternatively, clutter, feature extraction and target detection methods. Typical
jammers and targets can be transformed to Doppler-angle features differentiating a target and clutter or jammers are the
domains for target detection based on their different features. received signal shape in the transform domain. We will define
The features of clutter, jammers and target are shown in Fig. 3. the feature based on serval concepts firstly developed by
The ground clutter for airborne radar in angle-Doppler domains ourselves. The target detection schemes designed using
is extended in a tilted direction since the clutter Doppler artificial intelligence techniques include decision-tree based
frequency changes with its angle. Wideband jammer signals detector, feature parameter or vector based detector; and
are extended in Doppler domain and pointed in angle domain. machine-learning classifier. These detection schemes are
Target is pointed in both Doppler and angle domains. detailed as follows.
A. Target Detection Using a Decision Tree Similarly we can calculate the variances of the feature vectors
Target detection decision tree is a target-or-no-target for targets and clutter using training data. The standard
classifier of a hierarchical tree structure evaluating multiple deviations of the target and clutter feature vectors can be
features based on training data or analysis results using a expressed, respectively, as:
divide-and-conquer strategy. Fig. 5 is a decision tree for σ T = [σ t1 ,σ t 2 ,,σ tN ]T (4)
MTI/MTD radar target detection using the features in Fig. 1.
The feature data used for the detection are feature block sizes and σ C = [σ c1 ,σ c 2 ,,σ cN ] T
(5)
and their Doppler centers. We define the weighted distance between the data and the
Any feature
target as:
block with Doppler N

Yes
center >fT?
No
Dx ,T = ∑(x
i =1
i − µ ti ) 2 / σ ti2 (6)

Feature Same result if

block size Furthermore, the weighted distance between the data and
PRF changes?
<0.5? clutter is defined as:
Yes Yes No
No N

No
Target No
Feature
block size
Dx ,C = ∑(x
i =1
i − µ ci ) 2 / σ ti2 (7)
Target Target <0.5?
Yes Subsequently, target detection is based on the following
No
rule:
No If Dx ,T < Dx ,C , a target is declared for the data; otherwise,
Target
Target
no target is declared.
Fig. 5. A decision tree for feature-based target detection with MTI/MTD radars For this approach, we generally require that the feature
vector components are statistically uncorrelated.
In Fig. 5 target feature sizes are set to be between 1 and 3. If
the size of a feature is less than 0.5 with a large Doppler
frequency, it is generated from a random noise spike. Similar C. Target Detection Using Machine-Learning Classifiers
decision trees can be constructed for target detection with other Even more useful approach is to design a machine-learning
radar platforms or applications. This approach is simple, classifier using radar training data by establishing a hyperplane
straightforward and normally accurate. boundary between the target subspace and clutter subspace.
In order to construct a decision tree in Fig. 5, we need to use The boundary is formulated in the space with the feature vector
the following definition of feature block size. Feature block components we select for the classification of the radar data as
size is defined as the maximum distance between two any either target or interference. If the training data are available,
pixels in a feature block [4]. If a feature block contains only the boundary can be formulated by using an artificial neural
one pixel, the size is defined as zero. The feature blocks are network [6] or a support vector machine (SVM) [7]. The
generated from the thresheld image data in Fig. 4 using a advantage of this approach is that the different feature vector
region growing algorithm [5]. components do not need to be independent.

B. Target Detection Based on Feature Vector Distances IV. PRELIMINARY RESULTS

Another useful target detection approach is based on the We have achieved some preliminary target detection results
distance between the data feature vector and the mean feature using the proposed artificial intelligence approaches. We
vector of typical targets. Assume we use the feature vector x applied the decision-tree approach and the one using the
containing N components, which are likely to be feature block feature vector distances to the simulated radar data with some
size, Doppler frequency of the block center, Doppler spread, very promising results obtained.
range spread, and etc.
A. Target Detection Results Using A Decision Tree for A
x = [ x1 , x2 ,, x N ]T (1) MTI Radar
For typical radar targets and clutter, using training data we can We simulated aground MTI radar in C- band with 6000
calculate the mean feature vectors for target and clutter, datasets with half of them containing a moving target. The
respectively, as: waveform bandwidth is 1MHz, the radar PRF is 5kHz, the
μ T = [ µ t1 , µ t 2 ,  , µ tN ]T (2) number of coherent pulses during a beam dwelling is 16, the
SNR for received pulses is 0dB and CNR=40dB. The target
Doppler frequency is randomly selected between 0 and 5kHz
and μ C = [ µ c1 , µ c 2 ,  , µ cN ]T (3) for each CPI and the clutter Doppler frequency is zero. Using
the decision tree algorithm in Fig.5, the detection result is
displayed in Table I. Some misdetections of targets shown in
the Table are due to the fact that the Doppler frequency of
targets was randomly selected to be zero or close to zero. The only feature parameter used for target detection in this
application is the feature block size. We define the feature
TABLE I. CONFUSION MATRIX FOR CLUTTER/TARGET block size as the maximum size of any two image pixels inside
RECOGNITION the feature block.
Declared signal The feature vector distances generated from the four feature
Clutter Target blocks and the corresponding detection result are shown in
Actual Clutter 100% 0% Table II. From the table, we can find both target and clutter are
signal Target 0.2% 99.8% detected correctly based on the proposed algorithm.

Other than some target misdetections, the feature-based target

TABLE II. FEATURE BLOCKS GENERATED FROM FIG. 6 AND DETECTION
detection results using the decision tree were almost perfect. RESULTS
Feature Block B1 B2 B3 B4

B. Target Deteciton Results Using Feature Vector Distacnes Dx,T 52.2 85.6 4.2 58.9

For this approach, we used one feature vector with only one Dx,C 10.5 5.7 92.1 12.6
component, i.e., the feature block size. The radar data are
simulated from an airborne STAP radar with a 16-element Target? No No Yes No
array for beamforming processing. The target and clutter data
are generated in space-time domain, but their distinguishing
features are obtained in the angle-Doppler domain through
basis transformation. In the feature domain, the target and
REFERENCES
interference features, defined as feature blocks, must be
separated. Therefore, the feature blocks are either targets or
interferences such as clutter or jammers. The feature blocks [1] M. I. Skolnik, Introduction to Radar Systems, (3rd Edition.). New York:
McGraw-Hill, Inc., 2001.
for airborne radar generated using a region growing algorithm
and thresholding processing are shown Fig. 6(a, b). [2] D. C. Schleher, MTI and Pulsed Doppler Radar with Matlab (2nd
Edition), London: Artech House, 2010.
[3] J. Ward, “Space-time adaptive processing for airborne radar,” MIT
Lincoln Laboratory Technical Report 1015, December 1994.
[4] H. Deng, B. Himed, and M. C. Wicks, “Image feature-based space-time
processing for ground moving target detection,” IEEE Signal Processing
Letters, vol. 13, no. 4, pp. 216-219, April. 2006.
[5] S. A. Hojjatoleslami and J. Kittler, “Region growing: a new approach,”
IEEE Transactions on Image Processing, vol. 7, pp. 1079- 1084, July
1998.
[6] J. Schmidhuber, “Deep Learning in Neural Networks: An Overview,”
Neural Networks, vol. 61, pp. 85-117, 2105.
(a) (b) [7] C. Cortes and V. Vapnik, "Support-vector networks". Machine
Fig. 6. Feature blocks generated in Doppler-angle domain for airborne Learning., vol .20 , no. 3, pp. 273-297, 1995.
radar (a) 3-D plot; and (b) 2-D plot.