An Improved Deformable Part Model And Its Application In Pedestrian Detection

Deformable part models(DPM) is one of the best object detection algorithms according to VOC challenges. DPM has won the 2007 VOC challenge, since when it soon became the main detection algorithm for general object categories. The basic idea for DPM-based detection algorithm is dividing the target object category into subtypes, and splitting the target object into several parts. Correspondingly, it detects an object by detecting each part and detects an object category by detecting each subtype. On the foundation of a comprehensive and careful research on DPM, this paper proposes an improved DPM based detection algorithm for pedestrian detection, which has been demonstrated to be effective by substantial experiments. It’s supposed that this algorithm also applies to other detection tasks.While learning and using the DPM-based object detection algorithm to various detection tasks, we have found the high detection accuracy of DPM is at the expense of great computation burden, thus limiting the development of its application. This essay is focused on solving the problem of the low detection speed of DPM. We have put forward the single-point calculation model and the fast level location algorithm. Finally we present the Greedy DPM algorithm based on them.The main work and the innovations of this essay:Presents a comprehensive and careful research on the DPM based object detection theory and its implementations.Proposes the single-point calculation model which is able to limit the calculation of DPM score to a local point, making it possible to test local image positions if there is a target.Proposes the fast level locating algorithm which is able to locate the object’s level in the feature pyramid efficiently, avoiding the enormous computation of brutal search.Proposes the Greedy DPM algorithm based on single-point model and fast level locating technique. While achieving great acceleration, the greedy DPM merely requires adding another insensitive threshold to the original DPM, thus it’s easy to tune the parameters in face of complex situations.