ASFF-YOLOv5: Multielement Detection Method for Road Traffic in UAV Images Based on Multiscale Feature Fusion

• Published in: Remote sensing (Basel, Switzerland) Vol. 14; no. 14; p. 3498
• Main Authors: Qiu, Mulan, Huang, Liang, Tang, Bo-Hui
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.07.2022
• Subjects: Accuracy; adaptively spatial feature fusion; Algorithms; basic traffic geographic information database; Clustering; Datasets; Experiments; Feature extraction; Geography; Information retrieval; Neural networks; object detection; Object recognition; Occlusion; Receptive field; Remote sensing; road traffic multiple elements; Roads; Roads & highways; Scale invariance; spatial pyramid pooling fast; Statistical analysis; Traffic; Traffic information; Unmanned aerial vehicles; Vehicles
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs14143498

Road traffic elements are important components of roads and the main elements of structuring basic traffic geographic information databases. However, the following problems still exist in the detection and recognition of road traffic elements: dense elements, poor detection effect of multi-scale objects, and small objects being easily affected by occlusion factors. Therefore, an adaptive spatial feature fusion (ASFF) YOLOv5 network (ASFF-YOLOv5) was proposed for the automatic recognition and detection of multiple multiscale road traffic elements. First, the K-means++ algorithm was used to make clustering statistics on the range of multiscale road traffic elements, and the size of the candidate box suitable for the dataset was obtained. Then, a spatial pyramid pooling fast (SPPF) structure was used to improve the classification accuracy and speed while achieving richer feature information extraction. An ASFF strategy based on a receptive field block (RFB) was proposed to improve the feature scale invariance and enhance the detection effect of small objects. Finally, the experimental effect was evaluated by calculating the mean average precision (mAP). Experimental results showed that the mAP value of the proposed method was 93.1%, which is 19.2% higher than that of the original YOLOv5 model.