In recent years, many deep learning-based approaches for object detection in autonomous driving systems have been proposed and achieved great achievements. These approaches can be divided into two groups: onestage methods and two-stage methods. Compared with one-stage methods, two-stage methods achieve better detection performance. However, the performance of two-stage methods is still limited with objects at different scales and heavy occluded objects because there is little discriminative feature to exploit in the last convolution layer of the base network in two-stage architecture. To solve this problem, this paper proposes an improved two-stage framework based on Faster R-CNN architecture for object detection in autonomous driving systems. In the proposed framework, the base network based on VGG16 architecture is first adopted to generate the base convolution layers. To increase the performance of detecting objects at different scales, multi-feature concatenation modules are used at different convolution blocks of the base network. The proposed multi-feature concatenation modules combine all sub-layers of each convolution block to generate enhanced feature maps, which contain more discriminative features. All enhanced feature maps generated by multi-feature concatenation modules are then fed to an improved multi-layer region proposal network module. Each improved RPN contains a 1×1 convolution layer for compressing the input channel and a 3×3 dilated convolution layer for increasing the receptive field. To solve the issue of heavy duplicate proposals in traffic scene images, soft-NMS algorithm is adopted to keep final proposals. Finally, all good proposals are fed to a detection sub-network which includes a RoI pooling layer and fully connected layers for classifying objects and regressing the coordinates of each detected object. Experimental results on Pascal VOC dataset and KITTI dataset show that the proposed method outperforms Faster R-CNN in detection accuracy.