YOLO-ELWNet: A lightweight object detection network

Abstract

This paper proposes a YOLO-based efficient lightweight network (YOLO-ELWNet) for onboard object detection based on the YOLOv3. A channel split and shuffle with coordinate attention module is developed in the backbone block, which effectively reduces the size of model parameters and computational cost while maintaining the detection accuracy. A new feature fusion network is proposed in the neck block, where a cross-stage partial with efficient bottleneck module is put forward to improve the feature extraction ability and reduce the computational cost. The Scylla intersection over union-based loss function is utilized in the head block, which accelerates the convergence speed of the YOLO-ELWNet. The effectiveness of the proposed YOLO-ELWNet is validated on the open source KITTI vision benchmark. The performance of YOLO-ELWNet is superior to some mainstream lightweight object detection models in terms of detection accuracy and computational cost, which demonstrates its applicability for resource-constrained onboard object detection.