High-performance Object Detection and Tracking Using Deep Learning

Abstract

Human detection and tracking are two fundamental problems in computer vision, which have been cornerstones for many real-world applications such as video surveillance, intelligent transportation systems and autonomous driving. Benefiting from deep learning technologies such as convolutional neural networks, modern object detectors and trackers have been achieving much improved accuracy on public benchmarks. In this work, we aim to improve deep learning based human detection. Our main idea is to exploit semantic context information for human detection by using deeply learned semantic features provided by semantic segmentation masks. These segmentation masks play as an attention mechanism and enforce the detectors to focus on the image regions where potential object candidates are likely to appear. Furthermore, after reviewing some widely used detection benchmarks, we found that the annotation quality for small and crowd objects does not meet to a satisfied standard. Hence, we introduce a new dataset which includes more than 8000 images for detecting small and crowd targets in fixed angle videos. Meanwhile, a baseline detector was proposed to exploit motion channel features for boosting the detection performance. The experimental results show that our proposed approach significantly improve the detection accuracy for the baseline detectors. In addition to a novel method for object tracking, we propose to transfer the deep feature which is learned originally for image classification to the visual tracking domain. The domain adaptation is achieved via some “grafted” auxiliary networks which are trained by regressing the object location in tracking frames. Moreover, the adaptation is also naturally used for introducing the objectness concept into visual tracking. This removes a long-standing target ambiguity in visual tracking tasks and we illustrate the empirical superiority of the more well-defined task. We also experimentally demonstrate the effectiveness of our proposed tracker on two widely used benchmarks