Unattended ground sensors (UGS) are widely used in industrial monitoring and military operations. Such UGS systems are usually lightweight devices that automatically monitor the local activities in-situ, and transfer target detection and classification reports to the processing center at a higher level of hierarchy. Commercially available UGS systems make use of multiple sensing modalities (e.g., acoustic, seismic, passive infrared, magnetic, electrostatic, and video). Efficacy of UGS systems is often limited by high false alarm rates because the onboard data processing algorithms may not be able to correctly discriminate different types of targets (e.g., humans from animals). For example, discriminating human footstep signals from other targets and noise sources is a challenging task, because the signal-to-noise ratio (SNR) of footsteps decreases rapidly with the distance between the sensor and the pedestrian.
Seismic sensors are widely used for personnel detection, because they are relatively less sensitive to Doppler effects and environment variations, as compared to acoustic sensors. Current personnel detection methods, based on seismic signals, are classified into three categories: time domain, frequency domain, and time-frequency domain. Generally, time-domain analysis may not be able to detect targets very accurately because of the interfering noise, complicated signal waveforms, and variations of the terrain.
Although passive infrared (PIR) sensors have been used for detection and localization of moving targets, similar efforts for target classification have not been reported. This work makes use of a wavelet-based feature extraction method, called Symbolic Dynamic Filtering (SDF). The SDF-based feature extraction algorithm mitigates the noise by using wavelet analysis, captures the essential signatures of the original signals in the time-frequency domain, and generates robust low-dimensional feature vectors for pattern classification. The objective is to detect and classify different targets, where seismic and PIR sensors are used to capture the characteristic signatures. For example, in the movement of a human or an animal across the ground, oscillatory motions of the body appendages provide the respective characteristic signatures.
Seismic and PIR sensor data were collected on multiple days from test fields on a wash (the dry bed of an intermittent creek) and at a choke point (a place where the targets are forced to go due to terrain difficulties). During multiple field tests, sensor data were collected for several scenarios that consisted of targets walking along an approximately 150-meter-long trail, and returning along the same trail to the starting point.
The targets consisted of humans (male and female), animals (donkeys, mules, and horses), and all-terrain vehicles (ATVs). The humans walked alone and in groups with and without backpacks; the animals were led by their human handlers, and they made runs with and without payloads; and ATVs moved at different speeds (5 mph and 10 mph). There were three sensor sites, each equipped with seismic and PIR sensors. The seismic sensors (geophones) were buried approximately 15 cm deep underneath the soil surface, and the PIR sensors were collocated with the respective seismic sensors. All targets passed by the sensor sites at a distance of approximately 5 m. Signals from both sensors were acquired at a sampling frequency of 10 kHz.
In order to test the capability of the proposed algorithm for target detection, another data set was collected with no target present. The problem of target detection is then formulated as a binary pattern classification, where no target present corresponds to one class, and target present (i.e., human, vehicle or animal) corresponds to the other class. The data sets, collected by the channel of seismic sensors that are orthogonal to the ground surface and the PIR sensors that are collocated with the seismic sensors, are used for target detection and classification. For computational efficiency, the data were down-sampled by a factor of 10 with no apparent loss of information.
Similar with the movement type identification shown above, the target payload information can also be derived by performing another binary classification for both animal and human targets. The feature vectors extracted by SDF has large inter-class separation while small intra-class variance, and yet the intraclass differences between the with-payload and without-payload cases are still distinguishable.
Seismic and PIR sensors have their own advantages and disadvantages for target detection and classification. The seismic sensor is omni-directional and has a long range of detection, whereas a PIR sensor has a limited field of view (less than 180 ), which restricts the sensor from detecting target moving behind it. The seismic sensor is not site-independent and is vulnerable to variations in sensor sites, whereas a PIR sensor merely passively accepts the incoming infrared radiation and is independent of the sensor site. In order to improve the detection and classification accuracy while reducing the false alarm rate, it is recommended that the seismic and PIR sensor should be used together to provide complementary information to each other. Information fusion techniques are needed to combine the outputs of the two sensing modalities.
This work was done by Thyagaraju Damarla of the Army Research Laboratory; Xin Jin, Asok Ray, and Soumalya Sarkar of Pennsylvania State University; and Shalabh Gupta of the University of Connecticut. ARL-0147
This Brief includes a Technical Support Package (TSP).
Target Detection and Classification Using Seismic and PIR Sensors
(reference ARL-0147) is currently available for download from the TSP library.
Don't have an account? Sign up here.