To enhance the accuracy of environmental perception and object detection for intelligent vehicles, this paper proposes a novel feature fusion-based 3D object detection method. First, the Transformer model is improved by leveraging a multi-head self-attention mechanism to fully capture local spatial feature information while preserving diverse feature weights, enabling the extraction of point cloud and image features. An Unscented Kalman Filter (UKF) is then utilized to design a multimodal fusion system that combines image and point cloud data, ultimately achieving 3D object detection. The proposed model is trained and tested on both public and private datasets and is compared against several state-of-the-art algorithms. Experimental results demonstrate that the proposed feature fusion-based 3D object detection model achieves an improvement in mean Average Precision (mAP) of 0.34 percentage points and 3.03 percentage points compared to widely used algorithms such as F-ConvNet and F-PointNet, respectively. For vehicle and cyclist detection tasks, the mAP increases by 2.52 percentage points and 9.32 percentage points, respectively. Moreover, the inference results on real-world vehicle data align closely with the training evaluation outcomes. These findings indicate that the proposed algorithm improves the accuracy of object detection in perception systems to a certain extent and provides a valuable reference for the design of object detection systems in autonomous driving.