Target image tracking method based on deep reinforcement learning and space-time context

Abstract

The invention discloses a target image tracking method based on deep reinforcement learning and space-time context. The target image tracking method comprises the following steps: 1), obtaining an image xt from an input sequence through employing a feature extraction network at each time step t, and enabling the image xt to serve as a visual feature; enabling the visual features to pass through anSTC and a recurrent neural network, then extracting spatial-temporal features ct and a hidden layer state ht from the STC and the recurrent neural network respectively, and taking the spatial-temporal features ct as reference standards; 2) establishing a model; 3) carrying out model training; and 4) performing target tracking according to the prediction position of the model. The target image tracking method and the model provided by the invention have relatively high success rate and precision score in the tracking process, and also reflect that the DRST model based on reinforcement learningand space-time context provided by the invention can realize long-term tracking of the target object, so that tracking drift in the tracking process is avoided.

Description

technical field [0001] The invention relates to image processing, in particular to a target image tracking method based on deep reinforcement learning and spatio-temporal context. Background technique [0002] Different from the successful application of deep learning in visual fields such as target detection and target recognition, deep learning has many difficulties in the field of target tracking. The main problem lies in the lack of training data: deep learning models can effectively learn a large number of labeled training data, but the target Tracking only provides the bounding-box of the first frame as training data, so it is very difficult to train a deep model from scratch for the current target at the beginning of tracking. Aiming at the problems of tracking drift and slow tracking speed in the target tracking task, the present invention proposes a target tracking model (DRST) based on deep reinforcement learning (Reinforcement Learning) and spatio-temporal context...

AI Technical Summary

Problems solved by technology

Aiming at the problems of tracking drift and slow tracking speed in the target tracking task, the present invention proposes a target tracking model (DRST) based on deep reinforcement learning (Reinforcement Learning) and spatio-temporal context STC (Spatio-Temporal Context) learning

Images