A Novel Multi-Modal Deep Learning Approach for Real-Time Live Event Detection Using Video and Audio Signals
R Pavadareni, A. Prasina, V. Samuthira Pandi, Ibrahim Mohammad Khrais, Alok Jain, Karthikeyan · International Journal of Advanced Computer Science and Applications · 2025
Recent developments in live event detection have primarily focused on single-modal systems, where most applications are based on audio signals. Such methods normally rely on classification approaches involving the Mel-spectrogram. Single-modal systems, though effective in some applications, suffer from severe disadvantages in capturing the complexities of a real-world event, which thereby reduces their reliability in dynamically changing environments. This research study presents a novel multi-modal deep learning approach that combines audio and visual signals in order to enhance the accuracy and robustness of live event detection. The innovation lies in the use of two-stream LSTM pipelines, allowing for temporally consistent modeling of both input modalities while keeping a real-time processing pace through feature-level fusion. Unlike many of the recent transformer models, we are utilizing proven techniques (MFCC, 2D CNN, ResNet and LSTM) in a latency-aware and deployment-friendly architecture suitable for embedded and edge-level event detection. The AVE (Audio Video Events) dataset, consisting of 28 categories, has been used. For the visual modality, video frames undergo feature extraction through a 2D CNN ResNet and temporal analysis through an LSTM. Simultaneously, the audio modality employs MFCC (Mel Frequency Cepstral Coefficients) for feature extraction and LSTM to capture temporal dependencies. The features extracted from both audio and video modalities are concatenated for fusion. The proposed integration leverages the complementary nature of audio and visual inputs to create a more comprehensive framework. The outcome yields 85.19% accuracy in audio and video-based events due to the effective fusion of spatial and temporal cues from diverse modalities, outperforming single-modal baselines (audio-only or video-only models).