Cesur Karabacak

Knowledge Exploitation for Human Micro-Doppler Classification

Micro-Doppler radar signatures have great potential for classifying pedestrians and animals, as well as their motion pattern, in a variety of surveillance applications. Due to the many degrees of freedom involved, real data need to be complemented with accurate simulated radar data to be able to successfully design and test radar signal processing algorithms. In many cases, the ability to collect real data is limited by monetary and practical considerations, whereas in a simulated environment, any desired scenario may be generated. Motion capture (MOCAP) has been used in several works to simulate the human micro-Doppler signature measured by radar; however, validation of the approach has only been done based on visual comparisons of micro-Doppler signatures. This work validates and, more importantly, extends the exploitation of MOCAP data not just to simulate micro-Doppler signatures but also to use the simulated signatures as a source of a priori knowledge to improve the classification performance of real radar data, particularly in the case when the total amount of data is small.

Multi-aspect angle classification of human radar signatures

The human micro-Doppler signature is a unique signature caused by the time-varying motion of each point on the human body, which can be used to discriminate humans from other targets exhibiting micro-Doppler, such as vehicles, tanks, helicopters, and even other animals. Classification of targets based on micro-Doppler generally involves joint timefrequency analysis of the radar return coupled with extraction of features that may be used to identify the target. Although many techniques have been investigated, including artificial neural networks and support vector machines, almost all suffer a drastic drop in classification performance as the aspect angle of human motion relative to the radar increases. This paper focuses on the use of radar networks to obtain multi-aspect angle data and thereby ameliorate the dependence of classification performance on aspect angle. Knowledge of human walking kinematics is exploited to generate a fuse spectrogram that incorporates estimates of model parameters obtained from each radar in the network. It is shown that the fused spectrogram better approximates the truly underlying motion of the target observed as compared with spectrograms generated from individual nodes.

Radar simulation of human micro-Doppler signature from video motion capture data

The availability of data sets on which signal processing techniques may be tested is critical to the development of human detection, identification, and classification algorithms. However, in many cases real radar data of the desired characteristics may be expensive or difficult to obtain. In this case, synthetic or simulated data is desired. Much of the simulated data used in publications is derived from the Boulic kinematic model. But, the Boulic model is only valid for walking and is not applicable to compute the micro-Doppler signatures of other human motions. The Carnegie Mellon University motion capture library includes data from a wide range of human activities and provides the time-varying position of body parts. In this work, this video motion capture data is used to generate the radar micro-Doppler signature for many human activities. Observations about the micro-Doppler signatures computed are also shared.

Simulation of human micro-Doppler signatures with Kinect sensor

The availability and access to real radar data collected for targets with a desired characteristic is often limited by monetary and practical resources, especially in the case of airborne radar. In such cases, the generation of accurate simulated radar data is critical to the successful design and testing of radar signal processing algorithms. In the case of human micro-Doppler research, simulations of the expected target signature are required for a wide parameter space, including height, weight, gender, range, angle and waveform. The applicability of kinematic models is limited to just walking, while the use of motion capture databases is restricted to the test subjects and scenarios recorded by a third-party. To enable the simulation of human micro-Doppler signatures at will, this work exploits the inexpensive Kinect sensor to generate human spectrograms of any motion and for any subject from skeleton tracking data. The simulated spectrograms generated are statistically compared with those generated from high quality motion capture data. It is shown that the Kinect spectrograms are of sufficient quality to be used in simulation and classification of human micro-Doppler.

Feature selection for classification of human micro-Doppler

Dozens of features have been proposed for the use in a variety of human micro-Doppler classification problems, such as activity classification, target identification, and arm swing detection. However, the issues of how many features are truly required, which features should be selected, and whether or how this selection will vary depending upon human activity has not yet been rigorously addressed in the context of human micro-Doppler analysis. Moreover, most classification results are present for the case when the human directly walks towards or away from the radar. As the aspect angle between target and antenna increases, the observed micro-Doppler spread diminishes, leading to increasingly poor feature estimates. Thus, there is also a question of how features should be selected by taking into consideration estimate quality. This work examines the application of information theory to shed light on these questions. Mutual information is used to compute the contribution of features as a function of physical relevance and estimate quality. An importance ranking of features is derived, with results shown for arm swing detection and discrimination of walking from running.

Human activity recognition using a low cost, COTS radar network

Wireless sensor networks have been a subject of much interest as a means for wide area surveillance. Typically, sensors such as acoustic, seismic, infrared, magnetic, and ultrasonic sensor have been employed to date. Radar, although possessing important advantages such as being able to operate in all weather conditions and nighttime, has not much been used in these systems due to their high power requirements, high cost, and large size. Recently, however, low-cost, COTS radar nodes have been developed that enable their application as part of a wireless surveillance network. In this work, the BumbleBee Radar developed by Samraksh Company is used as part of a wireless radar network to monitor the activities of a human moving within the sensing region of the network. The human micro-Doppler signature measured by the BumbleBee radar is shown for a variety of activities and used as a basis for recognition. Various schemes for fusing sensor data are explored.

Radar simulation of different human activities via Kinect

The classification of different human activities with radar has been a widely researched topic in recent years. Oftentimes, when no experimental data is available, simulated data can be exploited to test classification algorithms. Kinematic models such as the Thalmann Model and motion capture (MOCAP) data are frequently used to simulate radar signatures of human movements. While the Thalmann Model provides a model only for human walking, MOCAP data has the capability to supply data for almost any type of human activity. However, most commercial MOCAP data acquisition systems are quite expensive, making it difficult to obtain MOCAP data. In this paper, economical, easily obtainable and practical Kinect sensor is used to develop a skeleton tracking algorithm. In this way, simulated radar micro-Doppler signatures for different people and activities are computed.

Automatic human activity classification using radar

Developing automatic target classification algorithms using radar is a widely researched topic in recent years. Among the targets classified in these algorithms, the most studied is humans. Classification of humans with high accuracy is very significant for many military and civil applications. Moreover, in these studies, besides the classification of human target, activities are also analyzed. Knowledge of the activity a person is engaged in can substantially change the alarm level in some applications. In this paper, an algorithm that automatically classifies walking, running, crawling, and creeping using radar is presented.

Indoor human activity recognition using BumbleBee radar

Human detection offers many advantages in applications of search and rescue, smart environments, and security. Infrared, acoustic, vibration/seismic and visual sensors have been often used in human detection and recognition systems. Radar offers unique advantages for sensing humans, such as remote operation during virtually all weather conditions, increased range, and better coverage. However, radar systems are typically very expensive and physically large. The BumbleBee radar, in contrast to most radars, is a low power pulse Doppler radar that is about the size of a business card. Moreover, it is a radar that can be integrated into indoor wireless sensor networks. In this work, the application of BumbleBee radar to human activity recognition by computing the human micro-Doppler signature is examined. Humans are complex targets that are capable of many motions. Every part of the human causes different reflection and every motion of the human has its unique micro-doppler signatures. The differences in micro-Doppler data of activities such as walking, running, and crawling that is gathered from low-cost, low-power radar is discussed.