A Cascade Framework for Unoccluded and Occluded Pedestrian Detection Aayush Ankit Dept. of Electronics Engineering Indian Institute of Technology (BHU), Varanasi aayush.ankit.ece11@iitbhu.ac.in Irfan Riaz Ahmad, Hyunchul Shin Dept. of Electronic and Communication Engineering Hanyang University ERICA irfancra@gmail.com, shin@hanyang.ac.kr Abstract— This paper presents a novel pedestrian detection framework for efficient detection of both unoccluded and occluded pedestrians, thereby proposing an efficient technique for pedestrian detection in real-time environment. Our framework consists of two layers of detection, the first layer using full body detectors for accurate detection of unoccluded pedestrians and then a cascaded layer of part based detectors to efficiently detect the occluded pedestrians. The full body detectors based techniques are state-of-the art for unoccluded pedestrian detection and the part based model is a viable choice for partially occluded pedestrian detection. In our part based model, we use six parts; three horizontal parts and three vertical parts thereby creating a model that is robust to varying degrees and types of occlusions. Each detection layer utilizes multiple modalities (cues) namely; intensity, dense stereo and dense flow. The use of part based detectors as the cascaded layer also increases the unoccluded pedestrian detection rate by correctly detecting the pedestrians that had been misclassified by the first layer. Thus, the second layer of part based detectors has a synergic effect on the first layer. Keywords- pedestrian; stereovision; occlusion; part-based methods I. INTRODUCTION Pedestrian Detection is very important and complex task in the field of Computer Vision, with significant implications in various practical applications such as video surveillance, content-based image/video retrieval, automotive sectors etc. However, detecting humans in images/videos is a very challenging task owing to the large variation in appearance, pose and clothing, as well as the varying backgrounds and environmental conditions makes the task arduous. The earlier pedestrian detection methods assumed full visibility of pedestrians in the scene. However, images from real environment scenes are replete with large number of partially occluded pedestrians as pedestrians move in the close proximity to other objects. Thus, in such cases the performance gets severely affected because the full body would not be visible then. The parts based approach for detection of partially occluded pedestrians is a viable solution as in this the decision will be based on the detection of unoccluded parts. In present times, the myriad discovery of cheap sensors for capturing stereo information and motion information enables the efficient exploitation of multiple modalities for pedestrian detection to enhance the detection results. In this paper, we present a cascaded two layered framework for efficient detection of both occluded and unoccluded pedestrians. The first layer consists of full body detectors based on multiple cues (intensity, dense stereo, dense optical flow) which detects the unoccluded pedestrians very accurately. The second layer is part based detectors cascaded with the first layer for accurate detection of partially occluded pedestrians. The part based detectors are trained to classify between the partially occluded pedestrians and non- pedestrians. The part based detectors are also trained on multiple cues. Multiple cues are used at both layers to utilise the information present in all the modalities. Since, the various modalities are uncorrelated; the fusion of the information extracted from each modality would help to make an ensemble classifier which combines the individual strengths of each modality classifier. The use of multiple cues results in significant improvement in pedestrian detection rate. In addition, the part based detectors also enhance the unoccluded pedestrian detection by correctly detecting the pedestrians that were misclassified by the first layer possibly because of complicated pose or orientation. This synergic effect between the first and second layers further improves the detection performance of unoccluded pedestrians. II. PREVIOUS WORK Pedestrian Detection has been a very popular research topic in the recent times because of the imperative need for intelligent vehicles due to large number of road accidents. Most surveys in this field are based on monocular pedestrian detection [1] [2] [3] [4] [5] [6] [7]. Most of the state-of-the-art systems discussed derive feature sets from intensity images and apply pattern classification methods on them. Non-adaptive Haar wavelet features were used in [7] [8] [9] and adaptive local receptive fields are proposed in [10]. Proceeding of the 2014 IEEE Students' Technology Symposium TS14IVC03 247 978-1-4799-2608-4/14/$31.00 ©2014 IEEE 62