An anchor-free detector and R-CNN integrated neural network architecture for environmental perception of urban roads

Abstract

Environmental perception of urban roads is a critical research goal in intelligent transportation technology and autonomous vehicles, and pedestrian location is key to many relevant algorithms. Because anchor-free detectors are faster and region-based convolutional neural networks have a higher accuracy in object detection and classification, we propose an integrated convolutional networking architecture combining an anchor-free detector with a region-based convolutional neural network in the environmental perception task. The proposed network achieves higher precision and increases inference speed by up to 30%. To acquire more accurate region boundaries than a coarse bounding box method, a semantic segmentation sub-network is adopted to predict an instance segmentation mask for each object, and more accurate segmentation results are obtained by using the Dice loss. Moreover, we present an assignment strategy using a modified feature pyramid structure and show that it improves mean average precision of pedestrian detection by 2% on average. Finally, we verify that the pretrained neural network is beneficial for small datasets. Overall, the results show that our model achieves higher precision than the approaches used for comparison.