Real-time Head Detection for Automated Passenger Counting in Embedded Systems

Head detection is a key problem for automated passenger counting systems. In recent decades, considerable effort has been expended to develop an accurate and reliable head detector. However, head detection is still a challenging task because of problems caused by variations in pose and occlusions. Recently, general object detection algorithms based on convolutional neural networks (CNNs), such as Faster R-CNN, SSD and YOLO, have been successful. However, these algorithms require the use of a Graphics Processing Unit (GPU) for real-time performance. In this study, we focused on developing real-time head detection in an embedded system. Starting with the Tiny-YOLOv3 network, we applied the following strategies to achieve real-time performance in a non-GPU environment. First, we reduced the input image size to 224x224. Second, we added an extra yolo layer to detect smaller heads. Third, we removed batch normalization. Finally, we conducted depthwise separable convolution rather than traditional convolution. Three public datasets, HollywoodHeads, SCUT_HEAD, and CrowdHuman, were exploited to train and test the proposed network, and Average Precision (AP) at Intersection over Unit (IoU) = 0.5 were used to evaluate the tests. Experimental results showed that the proposed network perform better and faster than Tiny-YOLOv3.