Object Tracking with Deep Learning and Swarm Intelligence

Abstract

Swarm Intelligence has been applied to object tracking in the recent decade. Despite the algorithm has consistently improved overtime, Swarm Intelligence based object trackers still perform poorly when evaluated against other state-on-the-art trackers on the publicly available online object tracking benchmark (TB50). Yet most existing research has not attempted to go beyond the algorithm to look for the cause of such poor performance. This thesis gives a full analysis on the shortcomings of Swarm Intelligence based algorithms and offers solutions to raise the accuracy and robustness of the Swarm Intelligence based object tracker. A novel weight adjustment function is first introduced to the Particle Swarm Optimisation algorithm for mitigating the problem of pre-mature convergence and loss of particle diversity. A Hybrid Gravitational Search Algorithm is then proposed to utilise information from all particles during the update process and to facilitate more thorough exploration within the search space. In addition, the Swarm Intelligence based algorithm only utilises spatial information from the current frame and fails to make use of temporal information such as past target location from previous frames. An object tracking framework is introduced to incorporate comprehensive target model prediction and occlusion handling into the object tracker. This is experimented to significantly improve tracking accuracy and robustness in the existence of very similar object in the video sequence. Furthermore, the impact of the feature representation is analysed in this thesis. The Stacked Denoising Autoencoder (SDAE) and Convolutional Neural Network (CNN) are tested and compared to the traditional colour histogram feature in Swarm Intelligence based object tracker. The seventh layer output of the Caffenet Model is found to generate a search space with significant reduction in the number of sub-optimums. This feature is further implemented and evaluated using 30 sequences from the OTB50. The change in feature almost double the accuracy and robustness of the Swarm Intelligence based object tracker, giving a comparable performance to STRUCK, the state-of-the-art in year 2013.