Accelerating Multi-target Visual Tracking on Smart Edge Devices

Abstract

\prefacesection{Abstract} Multi-object tracking (MOT) is a key building block in video analytics and finds extensive use in surveillance, search and rescue, and autonomous driving applications. Object detection, a crucial stage in MOT, dominates in the overall tracking inference time due to its reliance on Deep Neural Networks (DNNs). Despite the superior performance of cutting-edge object detectors, their extensive computational demands limit their real-time application on embedded devices that possess constrained processing capabilities. Hence, we aim to reduce the computational burdens of object detection while maintaining tracking performance.

As the first approach, we adapt frame resolutions to reduce computational complexity. During inference, frame resolutions can be tuned according to the complexity of visual scenes. We present DeepScale, a model-agnostic frame resolution selection approach that operates on top of existing fully convolutional network-based trackers. By analyzing the effect of frame resolution on detection performance, DeepScale strikes good trade-offs between detection accuracy and processing speed by adapting frame resolutions on-the-fly.

Our second approach focuses on enhancing the efficiency of a tracker by model adaptation. We introduce AttTrack to expedite tracking by interleaving the execution of object detectors of different model sizes in inference. A sophisticated network (teacher) runs for keyframes only while, for non-keyframe, knowledge is transferred from the teacher to a smaller network (student) to improve the latter’s performance.

Our third contribution involves exploiting temporal-spatial redundancies to enable real-time multi-camera tracking. We propose the MVSparse pipeline which consists of a central processing unit that aggregates information from multiple cameras (on an edge server or in the cloud) and distributed lightweight Reinforcement Learning (RL) agents running on individual cameras that predict the informative blocks in the current frame based on past frames on the same camera and detection results from other cameras.