Single Shot Multibox Detector (SSD) is the leading one-stage object detection method. However, the process of detecting objects at different scales with different levels of semantic information makes it difficult to fuse high-level features from deeper convolution layers with low-level features from shallower convolution layers. To mitigate the fusion process of different feature layers and create enhanced semantic information feature maps, this paper proposes a lightweight and efficient multi-feature fusion module. The proposed multi-feature fusion module includes concatenation operation to concatenate the features at different scales in a simple and efficient way, point-wise convolution layer to reduce the feature dimension, and bilinear interpolation to upsample the size of feature maps. Furthermore, ESPv2 network, a lightweight and efficient deep convolutional neural architecture, is adopted as the base network for generating base convolution layers from input image. The proposed multi-feature fusion module and base network make a significant improvement in both detection accuracy and inference speed. Experimental results on KITTI dataset show that the proposed model outperforms SSD in terms of detection accuracy while maintaining inference speed. In addition, an embedded system implemented on an NVIDIA Jetson TX2 is used for detecting cars in traffic scenes that shows the effectiveness of the proposed detector.