Towards High Accuracy Pedestrian Detection on Edge GPUs

Despite the rapid development of pedestrian detection algorithms, the balance between detection accuracy and efficiency is still far from being achieved due to edge GPUs (low computing power) limiting the parameters of the model. To address this issue, we propose the YOLOv4-TP-Tiny based on the YOLOv4 model, which mainly includes two modules, two-dimensional attention (TA) and pedestrian-based feature extraction (PFM). First, we integrate the TA mechanism into the backbone network, which increases the attention of the network to the visible area of pedestrians and improves the accuracy of pedestrian detection. Then, the PFM is used to replace the original spatial pyramid pooling (SPP) structure in the YOLOv4 to obtain the YOLOv4-TP algorithm, which can adapt to different sizes of people to obtain higher detection accuracy. To maintain detection speed, we replaced the normal convolution with a ghost network with a TA mechanism, resulting in more feature maps with fewer parameters. We constructed a one-way multi-scale feature fusion structure to replace the down-sampling process, thereby reducing network parameters to obtain the YOLOv4-TP-Tiny model. The experimental results show that the YOLOv4-TP-tiny has 58.3% AP and 31 FPS in the winder person pedestrian dataset. With the same hardware conditions and dataset, the AP of the YOLOv4-tiny is 55.9%, and the FPS is 29.