Design And Implementation Of A Visual Object Tracking System Based On Deep Learning

The application of deep learning has made great progress in artificial intelligence including computer vision.Object tracking,which has widespread applications in civil and military fields such as intelligent video surveillance,autonomous driving and reconnaissance of unmanned aerial vehicles,is an important research direction of computer vision.Unlike image classification and object detection,object tracking applies deep learning lately.Recently,some methods with top tracking accuracy prove that deep learning also has advantages in object tracking.However,these methods are far behind real-time and can only stay in the theoretical stage for now.Most people are still using the traditional method without the application of deep learning.In this paper,deep learning is introduced into the object tracking engineering.We design and implement an object tracking system leveraged by deep learning on NVIDIA Jetson TXl embedded platform.Our system offers three modes of operation,object tracking for local video,object tracking for real-time video,and object detection-tracking for real-time video.Object tracking method is the core of the whole system.Top-accuracy tracking methods are computationally intensive,while the Jetson TX1 platform has limited resources to perform computing.We can't blindly pursue high precision,and must make a trade-off between accuracy and speed.We draw on the idea of image comparison and regression network,design the network structure,use a large number of video sequences and images to train the network,and finally get the core tracking model of the system.The model,with 0.508 of AUC on OTB2013 dataset,has a higher accuracy score than other model which is widely used in engineering today.After implementing the core tracking method,we design and implement all modules in the system,including input module,output module,real-time detection module with 54.9mAP on Pascal VOC2007,and real-time tracking module based on the core tracking method.And then,we integrate all modules into a complete system.In the test,our embedded object tracking system achieved the expected results.The speed performance of 47 frames per second at the time of tracking meet the real-time requirement.And with that speed,it only had the power of 10.6w,which shows good energy efficiency.