CSP-Lite: Real-Time and Efficient Keypoint-Based Pedestrian Detection

Keypoint-based methods eliminate the need for anchor boxes and provide a simplified detection framework. Keypoint-based Center and Scale Prediction (CSP) achieves the state-of-the-art accuracy among pedestrian detectors. However, this accuracy corresponds to a high inference cost. To alleviate this problem and improve detection performance while ensuring speed, we propose a method called CSP-lite in this work. We propose a convolutional neural network trick based on epoch weights fusion, which improves network performance without additional training or inference cost, along with a simple and effective modified loss function. Additionally, we propose a highly efficient network module that extracts features more comprehensively. CSP-lite achieves 4.13$\%$ MR$^{-2}$ on the Caltech dataset with an inference time of only 6.3ms on RTX 2080Ti GPU. On the CityPersons dataset, it achieves 10.99$\%$ MR$^{-2}$ with an inference time of only 35.7ms on RTX 2080Ti GPU. The proposed method provides a balance between speed and performance, enhancing the practical application value of the method.