M-YOLO: A Nighttime Vehicle Detection Method Combining Mobilenet v2 and YOLO v3
Shan Huang, Ye He, Xiaoan Chen
Abstract
Abstract Vehicle detection at nighttime plays a vital role in reducing the incidence of night traffic accidents. In order to further improve the accuracy of nighttime vehicle detection, and to be suitable for constrained environments (such as: embedded devices in vehicles), this study proposes a deep neural network model called M-YOLO. First, M-YOLO’s feature extraction backbone network used the lightweight network MobileNet v2. Second, the K-means algorithm is reused to cluster the dataset to obtain the anchor boxes which are suitable for this paper. Third, M-YOLO uses the EIoU loss function to continuously optimize the model. The experiments showed that the average precision (AP) of proposed M-YOLO can reach to 94.96%. And ten frames per second (FPS) were processed in a constrained environment. Compared with YOLO v3, the proposed model performs better in detection accuracy and real-time performance.