A Tiny Pedestrian Detection Network Based On Deep Learning On Jetson TX2

As one of the most fundamental tasks in computer vision,object detection deals with classifying and locating instances of visual objects of a certain class in digital images.Object detection forms the basis of other computer vision tasks,such as video analysis and scene understanding.In recent years,the great progress of intelligent security and driverless vehicles,and the rapid development of deep learning techniques have brought new object detection algorithms and breakthroughs in the field of object detection,leading to more and more demands for fast and accurate object detection algorithms.Although great progress has been made in the field of object detection,challenges still exist in the following aspects: Detecting a small object with less feature and location information is difficult because of object rotation and scale changes.Speed up of high precision detection algorithms on high performance computer has long been a crucial task,let alone the embedded platforms.In this thesis,a tiny pedestrian detection algorithm is designed and a trade-off between detection precision and speed is achieved.The algorithm is successfully implemented on Jetson TX2(TX2),which is a kind of embedded software development environment,and it is in line with real needs.In summary,the contributions of this thesis are: A new tiny pedestrian dataset is obtained for the lack of tiny samples in common datasets.Based on YOLOv3-Tiny algorithm and the feature pyramid network struture,we integrate high and low layer feature maps together to get low receptive fields.Deep separable convolutional layers is introduced to replace traditional ones,so that a compact and lightweight convolutional neural network is constructed,which makes it more suitable for resource-constrained embeded platforms.At the same time,considering the hardware characteristics of TX2,change of memory addressing schemes and threading optimizations are also done to achieve real time detection performance.All experiments are tested on TX2.The experimental results show that the proposed algorithm has effectiveness on different datasets and achieves a good compromise between accuracy and speed.In addition,the proposed algorithm has simple structure and small weights,which has advantages on computational efficiency and portability.Also,with reasonable distribution of memory,we can make better use of TX2.