Vision-Based Human Behavior Recognition Using a Multiscale Convolutional Neural Network with Parallel Multi-Kernel Feature Fusion

Abstract: Automatic recognition of human behavior from video underpins applications ranging from surveillance and elderly-care monitoring to smart environments and human-computer interaction. Recognizing actions reliably is difficult because human movements vary in spatial scale, speed, viewpoint, and background clutter, and a single fixed receptive field rarely captures both fine gestures and coarse whole-body motion. This paper presents a vision-based behavior-recognition framework built on a multiscale convolutional neural network that extracts features through parallel convolutional branches with different kernel sizes and fuses them before classification. Video frames are preprocessed and localized to the human region, processed simultaneously at fine, mid, and coarse scales, and the fused representation is mapped to a behavior label with a confidence score. The recognition engine is implemented in Java with a deep-learning backend, while a Node.js layer provides a live monitoring interface. Evaluated against single- scale, two-scale, and recurrent baselines, the multiscale model attained an overall accuracy of about 94.2% with balanced per-class performance and a real-time throughput of roughly 45 frames per second. The principal contributions are a parallel multi-kernel feature-extraction design that captures complementary spatial granularities, a lightweight fusion strategy suited to real-time inference, and an integrated monitoring system that delivers interpretable, low-latency behavior predictions.

Keywords: Human behavior recognition; multiscale convolutional neural network; computer vision; action recognition; feature fusion; real-time inference; deep learning; video analytics.