HyMAD: A Hybrid Multi-Activity Detection Approach for Border Surveillance and Monitoring

Sriram Srinivasan, Srinivasan Aruchamy, Siva Ram Krishna Vadali · 2025

Read HyMAD: A Hybrid Multi-Activity Detection Approach for Border Surveillance and Monitoring by Sriram Srinivasan, Srinivasan Aruchamy, Siva Ram Krishna Vadali (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-14698.txt.

Introduction

Sriram Srinivasan, Srinivasan Aruchamy, and Siva Ram Krishna Vadali propose HyMAD, a system that watches borders using buried seismic sensors. Ground sensors are small and hard for intruders to spot, avoid, or damage, unlike cameras or fences. The hard part is telling apart events that happen at once, such as people walking, animals moving, and vehicles passing.

HyMAD is built to detect several activities at the same time from noisy ground signals. Missing that overlap can cause wrong labels, missed events, and a poor picture of what is happening. The hybrid design aims to separate the mixed signals well enough for reliable alerts. Stronger separation would make quiet, low-cost border monitoring more trustworthy.

Catalog record

CREATORSriram Srinivasan, Srinivasan Aruchamy, Siva Ram Krishna Vadali
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-14698

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

HyMAD: A Hybrid Multi-Activity Detection Approach for Border Surveillance and Monitoring Sriram Srinivasan, Srinivasan Aruchamy, Siva Ram Krishna Vadali Abstract Seismic sensing has emerged as a promising solution for border surveillance and monitoring; the seismic sensors that are often buried underground are small and cannot be noticed easily, making them difficult for intruders to detect, avoid, or vandalize. This significantly enhances their effectiveness compared to highly visible cameras or fences. However, accurately detecting and distinguishing between overlapping activities that are happening simultaneously, such as human intrusions, animal movements, and vehicle rumbling, remains a major challenge due to the complex and noisy nature of seismic signals. Correctly identifying simultaneous activities is critical because failing to separate them can lead to misclassification, missed detections, and an incomplete understanding of the situation, thereby reducing the reliability of surveillance systems. To tackle this problem, we propose HyMAD (Hybrid Multi-Activity Detection), a deep neural architecture based on spatio-temporal feature fusion. The framework integrates spectral features extracted with SincNet and temporal dependencies modeled by a recurrent neural network (RNN). In addition, HyMAD employs self-attention layers to strengthen intra-modal representations and a cross-modal fusion module to achieve robust multi-label classification of seismic events. e evaluate our approach on a dataset constructed from real-world field recordings collected in the context of border surveillance and monitoring, demonstrating its ability to generalize to complex, simultaneous activity scenarios involving humans, animals, and vehicles. Our method achieves competitive performance and offers a modular framework for extending seismic-based activity recognition in real-world security applications. I Introduction The use of seismic sensor networks has unlocked unprecedented opportunities for monitoring and interpreting activity patterns in various environments ranging from earth’s geological movement to border surveillance. Seismic signals, generated by ground vibrations, carry rich information about various sources, including anthropogenic activities (e.g., human footfalls, animal movements, vehicle passages). Accurate and automated detection and classification of these activities from seismic data are very important for border surveillance and monitoring. Despite significant advances in the classification of isolated activities like a lone person walking or a vehicle passing [ 1 ] [ 2 ] , the robust identification of multiple, simultaneously occurring activities like human and animal movements, or a vehicle passing alongside wildlife remains a central and largely open challenge in seismic signal analysis. Such overlapping signatures present a significant hurdle for traditional signal processing and machine learning approaches, as the combined vibrational patterns can obscure individual components, leading to misclassification or missed detections. Existing methodologies often struggle with the intricate fusion of diverse signal characteristics necessary to disentangle these complex, multi-source seismic events. Traditional seismic signal analysis has often relied on handcrafted spectral, temporal, or time-frequency features. However, these manually designed features typically struggle to capture the intricate and overlapping patterns that arise when multiple activities are happening simultaneously, which degrades performance and leads to poor generalizability in realistic, complex monitoring scenarios. Correctly identifying simultaneous activities is particularly important because missing one activity in the presence of another can significantly compromise the reliability of border surveillance systems. Prior studies have shown that overlapping seismic signals complicate detection and classification, often leading to incomplete or inaccurate monitoring records [ 3 ] , while seismic data analysis literature emphasizes that overlapping signals from natural and anthropogenic sources must be effectively decomposed to enable robust interpretation [ 4 ] . These findings highlight that resolving multi-source seismic events is not merely a technical improvement but a fundamental requirement for developing reliable security and monitoring systems. To overcome these limitations in multi-activity seismic event detection, we propose HyMAD (Hybrid Multi-Activity Detection) a novel deep neural framework designed to exploit the complementary strengths of both frequency and temporal features. Our HyMAD architecture initiates processing with a specialized SincNet-based convolutional layer that extracts interpretable, learnable frequency features via parameterized band-pass filters. These features serve as input to a recurrent neural network (RNN), which captures the spectral evolution and sequential dependencies. Both frequency and temporal representations are further refined through a self-attention module of their own, enabling each modality to contextually enhance its own features. Central to our approach is a cross-attention mechanism that adaptively fuses the two feature sequences, facilitating rich inter-domain interaction. The resulting fused representation is classified by a multi-layer perceptron (MLP) to identify one or more simultaneously occurring activities. The primary contributions of this paper are summarized as follows: • Novel HyMAD Architecture: We introduce a novel hybrid deep learning framework for robust detection of concurrent human, animal, and vehicle activities from raw seismic signals. • Principled Frequency Feature Extraction: Integration of a SincNet’s parametrized convolutional layer for learning interpretable and relevant frequency features directly from raw seismic waveforms. • Adaptive Multi-Modal Feature Fusion: A sophisticated attention-based strategy combining RNN-encoded temporal dynamics with SincNet-derived frequency features through self and cross-attention blocks for robust disentanglement of overlapping signatures. • Comprehensive Overlapping Activity Evaluation: Demonstrated superior performance in identifying complex, overlapping activity scenarios through extensive real-world seismic dataset experiments. II Related Work The field of seismic signal analysis for activity detection has evolved significantly, driven by advancements in sensor technology and machine learning. Our proposed Hybrid Multi-Activity Detection (HyMAD) architecture builds upon foundational work in seismic signal processing, deep learning for time series, and multi-modal feature fusion. This section summarizes existing research on recognizing human, animal, and vehicle activities, including their simultaneous occurrences, and discusses how our method addresses the shortcomings of previous approaches. II-A Traditional Seismic Activity Classification Early work on seismic activity classification aimed at distinguishing sources such as human footfalls, animal movements, and vehicle passages primarily relied on classical signal processing techniques. Methods such as Short-Time Fourier Transform (STFT) [ 5 ] , Wavelet Transform [ 6 ] , and various statistical features (e.g., Root Mean Square, Zero-Crossing Rate) were commonly employed to characterize seismic events. These extracted features were then typically fed into conventional machine learning classifiers like Support Vector Machines (SVMs) [ 7 ] , Random Forests [ 8 ] , or Hidden Markov Models (HMMs) [ 9 ] for activity classification. While these approaches offer a degree of interpretability due to their explicit feature engineering and can perform adequately for simple, isolated events in controlled environments, they are heavily reliant on extensive domain expertise for effective feature selection. This manual feature engineering process is often time-consuming and struggles to generalize across diverse environmental conditions or to novel activity types. More critically, these traditional methods demonstrate limited robustness and accuracy when faced with the inherent variability, noise, and particularly the overlapping signatures of multiple concurrent activities (e.g., human and animal, human and vehicle) prevalent in real-world seismic data. Their constrained capacity for adaptive feature learning and complex pattern recognition often leads to misclassification or missed detections when multiple activities co-occur and their unique features blend. II-B Deep Learning for Seismic Signal Analysis The advent of deep learning has significantly advanced time-series analysis, with increasing applications in seismic activity detection. Convolutional Neural Networks (CNNs) have been adopted for their ability to automatically learn hierarchical features directly from raw seismic waveforms, thereby reducing the reliance on manual feature engineering [ 10 ] . Recurrent Neural Networks (RNNs) [ 11 ] , including Long Short-Term Memory (LSTM) [ 12 ] and Gated Recurrent Unit (GRU) networks [ 13 ]

Read the complete text