Multi-sensing Pedestrian Detection Method Based on an Improved EPNet

Real-time and accurate detection of pedestrians is a necessary prerequisite for the safety of autonomous driving. To further improve the accuracy and performance of the pedestrian detection model, a multi-sensing pedestrian detection method for improving EPNet is proposed. Based on on-board laser radar and visual sensor capture pedestrian data, by analyzing the pedestrian detection task model misjudgment, for the automatic driving scene of positive and negative samples, with the aid of measurement learning and anchor-based algorithm of the multimodal learning model EPNet model structure improvement, and optimize the loss of the improved model function. The results show that the improved model anchor-based EPNet improves by 19.29% and 1.55% compared to EPNet precision and recall rate, respectively. This model shows significant advantages in target detection tasks based on multi-source heterogeneous pedestrian data and provides some reference for the application of other related models in pedestrian detection tasks.