ECDet: A Real-Time Vehicle Detection Network for CPU-Only Devices

Fei Gao, Zhiwen Wang, Jianwen Shao, Xinyang Dong, Libo Weng

Abstract: We study the problem of efficient vehicle detection for low-power devices in complex traffic environments and construct a lightweight real-time vehicle object detection model called Efficient Channel Detection (ECDet) to enhance the detection capability of vehicles at multiple scales. This work has three contributions: (1) the design of an efficient network backbone called Efficient Channel Network (ECNet); (2) the design of a lightweight real-time vehicle object detection model called ECDet; and (3) the design of a lightweight feature fusion structure called Lightweight Path Aggregation Network (LPAN). In addition, we quantitatively evaluated different methods on the UA-DETRAC dataset. The experiment shows that ECDet can quickly and accurately detect vehicles, and can be widely applied on embedded platforms such as intelligent transportation systems and unmanned driving systems.

External IDs:dblp:conf/icann/GaoWSDW23