Improvements to Tracking Pedestrians in Video Streams Using a Pre-trained Convolutional Neural Network

Target tracking has many applications in various fields. Millions of cameras are being used globally and people are constantly being watched everywhere. These cameras record over 48 hours of videos weekly which are impossible to be monitored manually. Many applications have been presented to improve the performance of pedestrian tracking. However, it still has remained a challenging topic. In this thesis, an automatic method is proposed for multiple pedestrian tracking. State-of-the-art detection has been combined with a tracking algorithm, followed by a novel post stage processing to increase the accuracy. Proposed automatic tracking system was compared with a state-of-the-art tracking algorithm which shows comparable accuracy when used with the original incomplete ground truth data. It is estimated to offer better accuracy with a more accurate ground truth data. The proposed algorithm offers potential improvements in both true positives as well as false negatives ratio when compared with the existing algorithm.