Research On Key Techniques Of Far-infrared Pedestrian Detection Based On Onboard Monocular Camera

Far-infrared(FIR)pedestrian detection based on onboard monocular camera has important academic significance and practical application value,which has currently become a research hotspot in computer vision.Compared with traditional pedestrian detection for video surveillance,main characteristics of onboard FIR pedestrian detection are the dynamism of the camera and pedestrians as well as the requirement of real-time detection.Currently,methods of onboard FIR pedestrian detection mainly include probabilistic template based methods and machine learning based methods.Because the machine learning based methods usually can obtain higher detection accuracy,it has become the mainstream method for onboard FIR pedestrian detection.Though lots of methods have been proposed for each stage of the machine learning based framework,robust and real-time FIR pedestrian detection is still a challenge technical problem.The reasons are as follows: A detection method must be able to handle dynamic and complicated traffic scenarios,a large length scale of pedestrian size and pedestrian appearance,the dynamic of the scene temperature,the lack of pedestrian detail information.In addition,the strict real-time requirement of the detection system leads to the fact that a detection method of too high complexity is not applicable.This dissertation focuses on the issue of pedestrian detection using onboard monocular FIR camera,which aims at solving robust and real-time pedestrian detection,and a systematic study of each key stage has been conducted for FIR pedestrian detection under the machine learning based framework.The major research contents consist of generation of FIR pedestrian Regions of Interest(ROIs),feature extraction of FIR pedestrian,classifier design of FIR pedestrian,scene contextual information extraction from FIR images.The main contribution of the dissertation is as follows:1)A FIR pedestrian segmentation method is proposed based on the road information and an adapted threshold to generate ROIs.Aiming at the current FIR pedestrian segmentation methods are easy to be affected by heat sources from the background,we propose a road horizontal line estimation algorithm to locate the Area of Interest(AOI)adaptively,and we perform horizontal projection based on the pixels' brightness to locate those strips where pedestrians may appear efficiently,so that the interference from the background heat sources could be reduced.On this basis,the wide of each estimated strip(denotes the wide of a pedestrian)is regarded as a key parameter to guide the execution of the subsequent segmentation algorithm to handle the diversity of the pedestrian scale.Experimental comparisons and analysis with the current excellent FIR pedestrian segmentation methods have verified the superiority and the effectiveness of the proposed method as well as the ability of resisting noise.2)We propose two kinds of local features to describe FIR pedestrians.Current FIR pedestrian detection methods which are only based on global features are usually of high computation load and they fail to handle the small inter-class variance between pedestrians and non-pedestrians.Aiming at this weakness,considering the extraction of local feature is usually of lower computation load and FIR pedestrians have some specific local patterns,we propose modelling the head and torso of a FIR pedestrian respectively to efficiently implement the classification of head and torso.Besides,we construct two kinds of pedestrian detection methods based on partial mode and global model in a parallel/cascaded manner according to the imaging characteristics of urban and rural scene based on the designed local features.Then,the effectiveness of these two kinds of local features has been verified by experiments.3)We propose a novel feature called Coded and Self-similarity Histogram of Local Intensity Differences(CSHLID)and two filters.Based on a machine learning based recognition framework,we propose a CSHLID feature to describe FIR pedestrians to ensure the global and local relationship of oriented gradients can describe the global and local pattern of pedestrians respectively,so that the description ability of local edge pattern and the global shape could be enhance.On this basis,in order to further improve the detection rate and reduce the false alarm rate,we propose a pedestrian head filter which fuses the brightness and gradient information and a road obstacle detection filter which is based on curve-fitting technique.Experiments show that the CSHLID feature can represent FIR pedestrians effectively,the filters can suppress false alarms efficiently and reduce the computation load of the subsequent machine-learning based classification,and the conducted pedestrian detection method based on the CSHLID feature and filters can work rather well under various scenarios(urban and suburban)or various seasons.4)We propose a novel method to build probabilistic templates for FIR pedestrians and we combinine the templates with a kind of heterogeneous features to implement pedestrian detection.We first build multiscale pedestrians' head probabilistic templates to release the large intra-class variance of pedestrian appearance according to the imaging characteristic of pedestrians' heads.Then,we choose the optimal template to perform pedestrian matching after the fusion of templates using an adaptive head location algorithm,so that the matching computation load can be decreased and the accuracy can be improved.Furthermore,heterogeneous features are obtained from those ROIs that has satisfied the template matching,and the features are fed to a statistical learning framework to implement the final pedestrian detection.Experiments show that this method has higher detection accuracy when compared with those methods which are only using probabilistic template based method or machine learning based method.5)We propose mining scene contextual information from FIR images and implement pedestrian detection based on scene context.Most of the current FIR pedestrian detection methods only focus on pedestrians' patterns themselves to perform classification,and we note that the scene context in the traffic environment could be helpful to the classification accuracy of ROIs.Accordingly,we propose three algorithms to mine the scene contextual information from the FIR images by studying the traffic scene context to aid the pedestrian detection.And then we integrate the contextual information(‘global' information)and the low-level information(‘local' information)from a framework of Bayesian maximum a-posterior.Accordingly,we propose a FIR pedestrian detection method using the scene context.Experiments on real-world datasets and benchmark datasets have demonstrated the effectiveness and prospect of the proposed method.