Study on Pedestrian Detection Based on an Improved YOLOv4 Algorithm

摘要

Pedestrian detection, which is widely used in automatic driving and pedestrian analysis, has always been a hot research topic in the fields of artificial intelligence and computer vision. With the development of deep learning, pedestrian detectors are becoming more accurate and faster. However, most of them can't strike a balance between accuracy and speed well. Therefore, in this study, we propose a pedestrian detection model based on an improved YOLOv4 algorithm which concerns both detection accuracy and efficiency. The detection model combines a new type of SPP (Spatial Pyramid Pooling) network and K-means clustering algorithm with YOLOv4 model for easier feature extraction. Furthermore, Mish activation function is applied in the neck of the detection model, replacing Leaky ReLU activation function to improve the detection performance. Our pedestrian detector achieves excellent results on the Caltech pedestrian dataset: 84.7% AP at a real-time speed of 36.4 FPS on Titan XP.