Pedestrian Detection In UAV Scene Based On Convolutional Neural Network

As one of important research topic in computer vision,pedestrian detection has always drawn much attention.The pedestrian detection has also evolved from the early image processing stage,the feature model classification stage to the deep learning stage gradually.Compared with traditional ground surveillance,the pedestrian detection from the low-altitude unmanned aerial vehicles(UAV)provides a "god perspective",and therefore facilitates deep understanding and analysis of pedestrian social behavior.Due to the arbitrariness of the UAV's view point and camera motion,pedestrian detection in UAV scenes has greater challenges,such as scale changes,camera shake.The computing power of the embed platform in UAV is also very limited,which poses additional challenges on the complexity of the detection algorithms.Based on the aforementioned discussions,this thesis attempts to study the pedestrian detection problem in the UAV scene from the following two aspects.First,this thesis analyzes and compares pedestrian detection techniques on the deep-learning-based framework,discusses the common object detection framework,and builds a pedestrian detection system based on SSD.In order to better adapt to the characteristics of pedestrians in the UAV scenes,we record a new dataset in many public places(such as schools,parks,playgrounds and squares),and manually label ground truths.Then,we propose a block strategy to train the SSD detection algorithm on the training set.and evaluate the fine-tuned detection models using the test set.The experimental results show that the detection models using the UAV data and block strategy get a good test result.Second,this thesis attempts to implement the designed pedestrian detection algorithm on the NVIDIA Jetson TX2 embedded platform.Considering the speed and portability of the SSD method,this thesis builds an embedded pedestrian detection system based on SSD,it consists of two complementary models: a robust global model and a fast local model.The former one carefully detects pedestrians on the entire images with a series of sliding windows;the latter one conducts local search around the previous detection results.The experimental results show that the trained SSD model could achieve effective and stable detection on the NVIDIA Jetson TX2 embedded platform,running at 5.3 frames per second.