Object tracking based on AI
Object tracking is an important research branch of computer vision, which has attracted great attention and been widely applied, such as surveillance of unmanned plane, automatic driving, inspection of pedestrians and cars, and so on. Since object tracking originated from the middle of last century, there are many researches have been accomplished. Actually, it is difficult to implemented for real-time object tracking in complex scene still. The object deformation, illumination variant, scale variant, fast motion and blur, and occlusion etc are greatly challenging for robust object tracking.
Currently, the main methods for object tracking can be typically categorized as the traditional one and the one based deep learning. The traditional methods perform well for real-time object tracking. However, due to limited feature extraction ability of traditional methods, the accuracy and robustness of these methods are strongly impaired by background changing. On the other hand, methods based on deep learning can obtain much more features for object representation, but they are usually time-consuming.
For real-time plane tracking in complex scene in practical project, this paper deeply delved into the research of multi-domain networks owing to its excellent accuracy, and proposed a concise object tracking framework, i.e. fast deep learning tracking networks such as Fig. 1, which efficiently solved the problems object pose variant, complex background clutters and so on through multi-layers’ features to enhance the object representation.
Fig. 1. FDLAT framework
Further, by optimizing the full-connection layer and regression layer, this method greatly promoted the tracking speed meanwhile it well improved the object recognition accuracy and object tracking precision. Therefore, the method achieved stable object tracking even with pose variant, background clutter, scale variant et. al. Simultaneously, this method is real-time such as the following Fig.2.
（a）object pose variant
Fig.2. Object tracking based FDLAT Net
About The Group
This research is from the group led by researcher Xu Zhiyong, Zhang Jianlin et. al., which dedicates to build intelligent information system with “acute eye” and “intelligent brain”. This group devotes to the wish that the machine can see the world and think about the world. The research of the group focuses on the AI (artificial intelligence) and extends it to related applications.
Fig. 3. Development of machine recognition
For the applications of future unmanned systems, the group has been carrying on the research data generation, object detection, object tracking, scene analysis and understanding, 3D reconstruction and pose estimation, object behavior estimation. In order to break through the spoon-feed training mode of current artificial intelligence, the group has been devoting to make the machine from being trained to studying by redefining and describing the intelligence, meta-learn (learning to learn). Further, the group try to drive the applications of these researches in many areas.
Simultaneously, the group has been dedicating to develop hybrid architecture high performance processing platforms for AI. These AI platforms have greatly supported the applications of these research, such as object detection, object recognition, image storage and image compression etc.
Fig. 4. High-speed processing platform
Fig. 5. Our proposed object tracking method comparison with state-of-art methods on public datasets
The group has published several research results on IEEE Access, Sensors, IET Im. Proc., IEEE Photon. J., etc. The Fig.5. illustrates the comparison of our object tracking method with the state-of-art methods.
Zhao Chunmei, Chen Zhongbi, Zhang Jianlin. Application of aircraft target tracking based on deep learning[J]. Opto-Electronic Engineering, 2019, 46(9): 180261.