HSG Algorithm for Pedestrian Detection

Until recently, the improvements in terms of the safety of cars have been focused on the reduction of the damages during the accidents. Based on this way of thinking, technologies like airbags, seat belts pretensioners and crash-smoothing features have been realized. Nowadays, in contrast of the previous trend, a big amount of work is done with the aim of try avoiding directly injuries. The meaning of this labour is to implement clever on-board systems with the capability of control the environment and send aproper alert message to the driver when a dangerous situations are incoming. The biggest difficulty aboutthe on-board systems development is due to the fact that chips have to capabilities to "discover" and"understand" the environment. These kind of systems are dubbed Advanced Driver Assistance Systems(ADAS). In ADAS World one of the most interesting type is called Pedestrian Protection System(PPS) . Its distinguishing feature is to detect both stationary and moving people inside a specic regionof interest (ROI), such as the vehicle perimeter and act specific actions if the collision is unavoidable. In this work was analysed the differences between two Feature detectors based on sliding windowapproach. The first one is called "Histogram of Oriented Gradients Detector". This kind of detector,introduced in 2005, thanks to N.Dalal and B.Triggs, is based on sliding window approach. One of themost important blocks is called Feature Descriptor, that represents an alternative representation of thereference image, with the aim of simplify the extraction of useful informations unwrapping the redundantones. The second kind of detector is called”Histogram of Significant Gradients”. Starting from the fact that HOG is promoted as superior to all other single features proposed for pedestrian detection, the purpose of many scientists was to merge HOG with more elaborated features in ”cascade schemes” in order to obtain higher detection rates. These ways of thinking, however, lead to very complex and power expensive hardware and software implementations.The main drawback born to the fact that HOG, in its implementation, requires inherently difficult floatingpoint operations and repeated memory accesses. The work proposed by M.Bilal and et., addresses these issues by adopting a fast and proficient object detection environment requiring a low-complexity feature based on Histogram of Gradients combined with a Lookup table (LUT)-based kernel SVM classifier. The detector, albeit being substantially less computationally hungry than the HOG, exhibits betterclassification when INRIA and ETH datasets are used. The proposed hardware implementation of HSG detector is the result of the instructions of M.Bilal andet.. This work has the purpose of propose many dedicated architectural solutions for dierent sectionsof the algorithm. This way of thinking leads to implement the structures on one side with a parallelapproach and on the other side looking at the power consumptions. Very interesting section is the "Tan Entity" one, because it speeds-up signicantly the performance ofthe overall system. The main drawback, in terms of delay and complexity, for the other kinds of hardwareimplementations is represented by the computation of the arctangent function. Various hardware implementations in literature uses complexarchitectures such as CORDIC to do it, but, in many cases the reached precision is worthless. In detail,the solution consists to solve a inequality, instead of the above mentioned trigonometric function. The nal section, called "Hist Block 8x8 Entity" has the aim of produce a vector of 36 elements, each ofthem represented on 4 bits. The amount of 36 elements is related to the fact that the entire 8 x 8 blockis divided in 4 blocks having dimensions 4x4. Every output block is called "histogram".