Deep Reinforcement Learning Based on Spatial-Temporal Context for IoT Video Sensors Object Tracking

The Internet of Things (IoT) is the upcoming one of the major networking technologies. Using the IoT, different items or devices can be allowed to continuously generate, obtain, and exchange information. The new video sensor network has gradually become a research hotspot in the field of wireless sensor network, and its rich perceptual information is more conducive to the realization of target positioning and tracking function. This paper presents a novel model for IoT video sensors object tracking via deep Reinforcement Learning (RL) algorithm and spatial-temporal context learning algorithm, which provides a tracking solution to directly predict the bounding box locations of the target at every successive frame in video surveillance. Crucially, this task is tackled in an end-to-end approach. Considering the tracking task can be processed as a sequential decision-making process and historical semantic coding that is highly relevant to future decision-making information. So a recurrent convolutional neural network is adopted acting as an agent in this model, with the important insight that it can interact with the video overtime. In order to maximize tracking performance and make a great use the continuous, inter-frame correlation in the long term, this paper harnesses the power of deep reinforcement learning (RL) algorithm. Specifically, Spatial-Temporal Context learning (STC) algorithm is added into our model to achieve its tracking performance more efficiently. The tracking model proposed above demonstrates good performance in an existing tracking benchmark.