Accelerating Multi-Object Tracking in Edge Computing Environment with Time-Spatial Optimization

Mengyang Liu,Anran Tang,Huitian Wang,Lin Shen,Yun-Han Chang,Guangxing Cai,Daheng Yin,Fang Dong,Wei Zhao
DOI: https://doi.org/10.1109/cbd54617.2021.00055
2021-01-01
Abstract:Multi-Object Tracking (MOT) task is a key issue in computer vision. It has capacious application prospects in survelliance, self-driving, and augmented reality. Benefit from the continuous progress of Deep Neural Network (DNN), the accuracy of the MOT algorithm has been significantly improved. Nevertheless, limited to computing power, achieving real-time DNN-based MOT is difficult in embedded systems. In reality, there are many wasteful and unnecessary computations in traditional frame-by-frame full-size video analysis. Therefore, in this paper, we propose a strategy that optimizing the execution of a traditional MOT pipeline in the dimension of time and space. In the temporal dimension, DNN only works in periodic keyframes while using a lightweight model for quickly generating results in the common frames. In the spatial dimension, we design an image density region discriminator to narrow down the input size of DNN. An edge device is introduced to perform end-edge collaborative computing to further accelerating the execution. Additionally, an end-edge parallel computing mechanism is designed that performing dynamic decisions based on the computing power and network environment between end and edge. Moreover, we rebuild the DNN model by TensorRT to optimize the model structure of DNN. By integrating the above approaches, the system can achieve 17.6 ~ 38.1 × speedup ratio, while with 3%~10.4% absolute tracking accuracy sacrifice and can be deployed in an unstable network environment.
What problem does this paper attempt to address?