To address the insufficient feature extraction capability of the PointRCNN object detection algorithm in complex scenes, we first design a local-global attention module (LGAM). LGAM computes attention weights from the local geometric relationships between each central point and its neighbors, enabling effective fusion of local features. Simultaneously, global contextual features are captured via a bilinear regularization method, and local and global features are then fused for collaborative optimization. Next, we introduce a multi-scale kernel convolutional attention module (MKCAM), which dynamically aggregates multi-scale features by parallelizing standard and dilated convolutions and incorporates a channel-pooled spatial attention mechanism. Both LGAM and MKCAM are cascaded into the original PointRCNN point-cloud encoding network to enhance its feature extraction capacity. Furthermore, to mitigate misdetections caused by the fixed IoU threshold in traditional non-maximum suppression (NMS), we propose a fuzzy NMS that adaptively assigns IoU thresholds based on object size and scene density. By integrating the improved point-cloud encoder with fuzzy NMS, we present an enhanced PointRCNN algorithm. Experimental results on the KITTI dataset show accuracy improvements of 1.05%, 3.43%, and 1.33% for cars, pedestrians, and cyclists, respectively. On our self-collected roadside dataset with sparse point clouds, detection accuracies for the three classes increased by 1.3%, 2.71%, and 2.9%, respectively, confirming the effectiveness and generalization ability of the proposed method.