Radar - Acoustic Vehicle Classification System based on shallow Convolutional Neural Network

Rapid urbanisation and growing traffic volumes have intensified the demand for efficient vehicle classification (VC) systems in Intelligent Transportation Systems (ITS). Traditional unimodal approaches, such as loop detectors or cameras, often fail under adverse conditions, limiting accuracy. Previous works relied mostly on single-modality sensors, which required complex algorithms to mitigate noise and cluttered environments. Many also adopted standard machine learning with hand-crafted features, which are difficult to design and sub-optimal. Moreover, limited research has validated multimodal frameworks such as dual-channel CNNs with real-world data. This study addresses these challenges by developing a multimodal shallow Convolutional Neural Network (SCNN) that integrates radar and acoustic sensors, exploiting their complementary strengths. Radar offers robustness in varying weather and lighting conditions, while acoustic sensors capture distinctive vehicle sound signatures from any angle. The objectives of this research are threefold: (i) to develop a dual-channel VC framework using radar and acoustic modalities, (ii) to design a dual-channel shallow CNN for efficient vehicle classification, and (iii) to validate the model using real-world traffic data. Experimental evaluation explored variations in time window lengths, spectrogram sizes, fusion stages, and operators. The proposed multimodal SCNN achieved a maximum classification accuracy of 96.7% with a 1-second time window, 128x 128 spectrogram, and late-fusion concatenation. In contrast, unimodal models achieved 89.4% (radar-only) and 91.2% (acoustic-only), confirming the benefit of multimodal fusion. Decision-level fusion consistently outperformed pixel-level fusion, with concatenation superior to summation. Compared with prior studies, which typically reported accuracies of 90-94% using unimodal sensors and conventional ML or deeper CNNs, the proposed approach not only achieved higher accuracy but also reduced computational complexity due to its shallow design. This shows that lightweight multimodal fusion networks can match or surpass state-of-the-art methods without the heavy resource demands of deeper models. The key contributions of this research are: (i) the design of a novel dual-channel shallow CNN for radar-acoustic vehicle classification, (ii) empirical validation of multimodal fusion for improved robustness and accuracy, and (iii) demonstration of superior performance relative to existing works using real-world data. Overall, this study delivers a reliable, precise, and efficient framework for next-generation ITS, enabling cost-effective, real­time vehicle classification for smarter urban traffic management.