Abstract
<title>Abstract</title> <p>Advanced Driver Assistance Systems (ADAS) have attracted massive interest recently, fueled by rapid breakthroughs in intelligent transportation and autonomous driving technologies. Accurately detecting surrounding objects is a critical requirement for ADAS, specifically under complex traffic environments where individual sensors frequently suffer from occlusion, noise, and limited sensing capability. To address these challenges, this paper introduces a Cross-Attention based Multi-Sensor Fusion framework for Robust Object Detection in Advanced Driver Assistance Systems (CAMSF-ADAS). The proposed model utilizes multimodal sensor data acquired from camera, LiDAR, and radar for robust perception and object recognition. Spatiotemporal calibration is performed to align camera, LiDAR, and radar data into a unified coordinate framework for enhancing spatial consistency and temporal synchronization among sensors. Besides, a Multi-Head Cross-Attention Fusion Module is introduced to efficiently integrate complementary information from different sensing modalities through dynamic feature interaction. For object detection, the YOLOv12 network is used to accurately localize and identify surrounding traffic objects under diverse driving scenarios. In addition, a Temporal Fusion Transformer (TFT) model is employed for classification to capture temporal dependencies and improve discriminative learning performance. To further improve the classification capability, the hyperparameters of the TFT model are optimized using the Improved Hippopotamus Optimization Algorithm, which enhances convergence behavior and search efficiency. Experimental analysis is carried out, and the obtained results show that the proposed CAMSF-ADAS model obtains superior performance compared to existing approaches. Therefore, the proposed model can be considered an effective solution for robust multi-sensor object detection in ADAS environments.</p>