University of Applied Sciences Mittweida and Chemnitz University of Technology at TRECVID 2018

The analysis of video footage regarding the identification of persons at defined locations or the detection of complex activities is still a challenging process. Nowadays, various (semi-)automated systems can be used to overcome different parts of these challenges. Object detection and their classification reach ever higher detection rates when making use of latest cutting-edge convolutional neural network frameworks. In our contribution to the Instance Search task, we specifically discuss the design of a heterogeneous system which increases the identification performance for the detection and localization of persons at predefined places by heuristically combining multiple state-of-the-art object detection and places classification frameworks. In our initial appearance to the task of Activity of Extended Video (ActEV) which is engaged in the detection of more complex activities of persons or objects, we also incorporate state-of-the-art neural network object detection and classification frameworks in order to extract bounding boxes of salient regions or objects that can be used for further processing. However, a basic tracking of objects detected by bounding boxes needs a special algorithmic or feature-driven treatment in order to include statistical correlations between the individual frames. Our approach describes a simple but yet powerful method to track objects across video frames.