| Pedestrian detection research has been a hot spot in computer vision for a longhistory. As the concept of intelligent transportation rising today, accurate pedestriandetection in road traffic is necessary for avoiding collision to save lives.Therefore, thestudy of real-time pedestrian detection technology is extremely valuable andpromising.Because of the non-rigidity of the target, the majority changes of the pose anddress plus the complexity of the scene changes, pedestrian detection is challenging inthe very beginning of constructing the feature descriptor. Recent years, the rise ofpedestrian detection methods based on statistical learning kind of solve the problem ofnear scale pedestrian detection..However, most effective algorithm has computationalchallenging, the can not reach the real-time performance.Besides, almost all methodhas a disastrous performance drop in the low resolution pedestrian case,and thisproblem even doesn’t get enough reserch attemtion by far.As a matter of fact,the lowresolution pedestrian detection is the keypoiont of avoiding traffic accident in the roadtraffic scene.For above reasons,I did a wide study in low resolution pedestriandetection and real-time performance. Based on vedio sequence which obtained by amonocular camera mounted on vehicle. I give the resolution of the two problemrespectively.1. A new detection algorithms. we did feature space transformation which issimilar to principal component analysis (PCA).We use the resolution awaretransformation matrix transform the feature vectors draw from different resolutionsamples to the same feature subspace to establishthe relationship betweenhigh-resolution and low-resolution pedestrian features.This improve the capability offeature descriptors in low resolution.thus reduce the false positive rate and missingrate in low-resolution case. In addition, we give parameters optimizing method tomeet the above idea. By splitting the non-convex quadratic optimization problem intotwo sub-convex optimization problem and do the optimization process respectively,we can get the optimizing model.2. A new ROIs segmentation method.We did it through geometric constraints.This approach is based on camera parameters calibration to determine the ransformation matrix between wold coordinates and image coordinates.We measurethe distance and pixels of pedestrian in the image on ground truth to determinedistance aware position and height in image.Then, we can give the scale awareROIs.This method can Cooperate with multi-threading and SSE technology togetherimprove the detection speed of single-frame images. By resizing the detector size andSubspace transformation, we can achieve a real-time mid-distance detection. |
