ESPPNet: An Efficient Progressive Spatial Pyramid Pooling Network for Real-Time Traffic Object Detection

Traffic object detection based on computer vision (CV) can usually be deployed on the embedded computing platform of autonomous vehicles or autonomous aerial vehicles (AAVs), to provide critical information about traffic scenes for autonomous driving or traffic management. However, due to limited computing resources, there is a need for small, lightweight, and reliable object detectors. As an emerging technology, spatial pyramid pooling methods have great potential in improving the detection performance of real-time object detectors. Most of the existing works focus on the development of more complex spatial pyramid pooling methods for higher accuracy, but real-time performance is also important in the ever-changing traffic scene. Thus, to balance the tradeoff between real-time detection and accuracy, we design a solution for real-time traffic object detection: a novel real-time object detector, named ESPPNet. Specifically, we propose an efficient plug-and-play spatial pyramid pooling method (ESPP). The method consists of a progressive spatial pyramid pool structure (PSPP) and a multi-scale feature enhancement module (MFEM). We first use PSPP to capture multi-scale feature maps with richer nonlinear features. Then, MFEM is used to establish effective long-range dependencies for multi-scale features. Experimental results on the VisDrone and SODA10M public datasets demonstrate that our method can achieve better real-time performance, less resource utilization, and higher accuracy, compared with other state-of-the-art methods. Note to Practitioners—This work aims to develop a practical and efficient object detection system for deployment in resource-constrained autonomous vehicles tasked with performing traffic object detection. Traffic object detection technology plays a crucial role in providing environmental perception capabilities for autonomous vehicles and autonomous aerial vehicles. It also serves as an essential means for traffic managers to acquire road information. However, existing object detectors generally fall short of capturing several crucial aspects, such as the interactions between different depth features or the independence of features of different receptive fields. This can lead to a bottleneck in feature extraction capabilities. To address these issues, our work proposes a plug-and-play lightweight module to support real-time traffic object detection on both mobile CPU and GPU, enabling the design of an efficient and practical end-to-end object detector. We have validated the performance of our proposed method on multiple traffic scene datasets.