Research on Embedded Multimodal Fusion Algorithms for Intelligent Perception

Multimodal fusion technologies play a crucial role in intelligent perception systems. Traditional fusion algorithms face challenges on embedded platforms due to limited computational resources, stringent real-time requirements, and strict power consumption constraints. This paper proposes an embedded multimodal fusion algorithm based on lightweight feature extraction and adaptive weight allocation. A depthwise separable convolutional network is constructed to efficiently extract visual features, while MFCC-based techniques are used to obtain audio modality information. An attention mechanism is adopted to dynamically adjust the weight of each modality. Experimental results on an ARM Cortex-A78 processor demonstrate that the proposed algorithm achieves a recognition accuracy of 92.8%, with an inference time of only 15.2 ms and a 34.6% reduction in power consumption. Compared with traditional methods, the proposed approach exhibits significant advantages, confirming its practicality and effectiveness in embedded intelligent perception systems.