Lightweight Neural Network Model and Algorithm for Pedestrian Detection

Traditional pedestrian detection methods have poor robustness. Deep learning-based methods have shown high performance in recent years but rely on substantial computational resources. Developing a lightweight, deep learning-based pedestrian detection algorithm is essential for applying deep learning-based algorithms in resource-limited scenarios, such as driverless and advanced driver assistance systems. In this article, an improved model based on YOLOv3 called “YOLOPD” (You Only Look Once—Pedestrian Detection), is proposed. It is obtained by constructing a self-attentive module, introducing a CIOU (Complete Intersection over Union) loss function and a depth separated convolutional layer. Experimental results show that on the INRIA (National Institute for Research in Computer Science and Automation), Caltech, and CityPerson pedestrian dataset, the MR (miss rate) of the model YOLOPD is better than that of the original YOLOv3 model, and the number of parameters is reduced by about 1/3, which significantly improves the speed of network derivation while improving detection accuracy.