Enhancing Pedestrian Detection Algorithm Based on YOLOv3-ES Network
Chenxu Li, Yiheng Sun · 2023
Pedestrian detection has wide applications but faces challenges such as overlapping pedestrians, bad weather conditions and small targets that strict its performance. To address these issues and enable the broader use of pedestrian detection, this paper proposes an improved detection algorithm called YOLOv3-ES based on YOLOv3. YOLOv3-ES incorporates ECA attention mechanism and SPP-net with YOLOv3. The ECA attention module operates on the three feature maps output by Darknet53, while the SPP-Net operates on the last feature map that has been processed by the ECA attention module. ECA attention mechanism can adaptively select useful features for detection, which helps to improve the detection rate and efficiency to solve the problem of low detection rate posed by overlapping pedestrians and bad weather. SPP-net can enlarge the receptive field through multi-scale maxpooling operations and fusion operations, which helps to improve detection accuracy, especially for small targets that are difficult to detect. Based on the INRIA dataset, the final results of the experiments show that YOLOv3-ES improves mA$P$by 3.67% over YOLOv3 and detects overlapping pedestrians and small targets better.