Pedestrian clothing detection, which is dedicated to the detection of pedestrian clothing, is of great significance in pose estimation, pedestrian classification, security and so on. We propose an algorithm named YOLOV3-PCD (An Improved YOLOV3 for Pedestrian Clothing Detection) for pedestrian clothing detection, which finds out the types and positions of clothes on pedestrians. Since there is no available dataset for pedestrian clothing detection task, we build our own dataset in which most objects are large and re-cluster the anchor boxes as well. In potential application fields of pedestrian clothing detection such as pose estimation and security scenarios, targets that need to be detected are usually large. So, we remove the scale used to detect small objects in the original YOLOV3. In addition, we simultaneously consider the detection of the rest medium and maximum scales, and introduce the down-sampling parallel to the original YOLOV3's up-sampling. Based on the above two improvements, we effectively increase the propagation and reuse of features, and improve network performance in the big object detection. In the end, for the application of embedded device in different scenarios, we prune the network to make it fast and small. Experiments show that the mAP of our proposed model reaches 91.99% which is 2% higher than the original YOLOV3 model and the number of parameters reduces to 28.74% of the original YOLOV3 model after pruning.
展开▼