Study Of Vehicle Detection Based On Deformable Part Models

Vision-based vehicle detection is an important issue in driving assistance system andcomputer vision. It aims at detecting vehicles appearing around the ego vehicle, so as to alertthe driver about the driving condition and possible collisions with other vehicles. Recently,the deformable part model (DPM) has been widely used in object detection for its highaccuracy and efficiency. Most of previous modified DPM methods use histograms of orientedgradients (HOG) as the feature descriptor. However, HOG limits the performance of DPM, asit cannot deal with noisy edges and ignore the flat areas while focusing on edge areas.Considering sparse coding, an increasingly popular and effective method for featurerepresentation, can obtain sparse codes which are robust to noise and can be used toreconstruct the input images through a learned dictionary, consequently this thesis proposes asparse coding based DPM for on-road multiple vehicle detection. The histograms of sparsecodes (HSC) computed with the dictionary learned by maximum correntropy criterion (MCC)are used as the features of DPM, instead of HOG, where the MCC is a local measurement thatcan make the features to be more insensitive and discriminatory. And an occlusion handlingmethod based on MCC is proposed to improve the accuracy of DPM. Compared with theoriginal DPM method, the proposed method has several advantages:(1) It can handle thenoises in input images uniformly within the correntropy framework during the reconstructionstep of sparse coding.(2) The sparse coding based feature vectors can represent moreinformation of image patches.(3) The nonlinear optimization problem in MCC-baseddictionary learning step can be simplified to a quadratic problem by using the half-quadraticoptimization algorithm to approximately maximize the objective function.(4) Occlusioninformation can be inferred from the proposed occlusion handling method.Compared with state-of-the-art approaches, the proposed method is more robust anddiscriminative, and improves the DPM method’s precision on the PASCAL VOC2007datasetand Toyota Motor Europe motorway dataset.