毕业生知识库

Object detection is an enduring topic in the field of computer vision. As a typical task of object detection, pedestrian detection has significant practical applications in the area of security, autonomous driving, intelligent surveillance and robotics applications. After more than ten years of exploration, pedestrian detection technologies have greatly developed and a series of classic algorithms have been proposed. Especially in recent years, with the introduction of the deep learning methods, the pedestrian detection models have gained surprising improvement. Despite deep learning based models can achieve good performance without designing hand-crafted features, low efficiency and high hardware requirement may limit their practical use. As a consequence, various deep network models which make compromises between speed and accuracy are proposed. Although excellent performance and efficiency has been achieved, pedestrian detection still has a long way to go for the following reasons. Firstly, the pedestrian objects in the road scenes are too small to contain enough feature points. However, most pedestrian detection frameworks perform poorly in detecting small targets. Secondly, the road traffic environment is complex and the pedestrian occlusion problem caused by various interference is serious. Thirdly, pedestrians always have many additional attributes, such as hats, bags, suitcases, etc, which increase the complexity of pedestrian detection.

 

In this paper, we study both traditional methods and deep learning models in pedestrian detection and integrate them to provide new solution for this task. Overall, the main contributions of this paper can be summarized as follows:
1. In this paper, an end-to-end multi-scale model based on Faster R-CNN is proposed to solve the problem of small target problem in Faster R-CNN framework. The model is free of external proposal detection and feature extraction methods and capable of learning and selecting different scales of feature information based on the size of the candidate boxes. The structure of our model is brief and easy to implement and extend.
2. The model is extended on the basis of Faster R-CNN. A neural cascade structure is proposed to solve the problem of hard negative examples mining and false positive samples. A strong classifier composed of multiple MLP (multi-layer perceptron) based weak classifiers is constructed and applied to judge the available candidate boxes in advance. Meanwhile, we train the network with selected hard examples, which relieves the computational burden and improve the accuracy as well.
3. A variety of network structures are explored and analyzed, and we conduct a series of experiments to evaluate the performance. The end-to-end multi-scale model presented in this paper is fast and performs well in the corresponding evaluation methods of certain open pedestrian detection data sets. Some crucial tricks used in experiments are summarized and analyzed as well.