The ultimate purpose of a pedestrian-detectionsystem (PDS) is to reduce pedestrian-vehicle-related injury. Mostsuch systems tend to adopt expensive sensors, such as infrareddevices, in expectation of better performance. In comparison, alow-cost optical-camera-based system has much potential practicalvalue, including a greater detection range, and can easily betrained to detect other objects. However, such low-cost systems aredifficult to design (e.g., little original information can be collected,and the scene is very complex). To address these problems, aneffective and reliable classifier is needed. The classifier should havea proper structure, its features need to be well selected, and a largenumber of high-quality samples are necessary for training. In thispaper, we present a low-cost PDS which only uses a single opticalcamera. We design a cascade classifier to achieve an effective andreliable detection. First, our system scans two sequential framesat each zoom scale with a sliding window. Second, with eachwindow, both appearance and motion features are extracted. Awell-trained cascade classifier, combining statistical learning witha decomposed support-vector-machine classifier, then determineswhether the window contains a human body. At the same time,to provide as much information as possible about the pedestrian,a small-scale weighted template tree trained by a coevolutionaryalgorithm is adopted to identify each pedestrian’s direction, andthe distance of each from the vehicle is also provided using anestimation algorithm. During the training procedure, we select keyfeatures by using the AdaBoost algorithm and a large numberof high-quality samples. Experimental results demonstrate thatthe system is suitable for pedestrian detection in city traffic: Thedetection speed is more than 10 ft/s, the detection rate reaches80%, and the false positive rate is no more than 0.3‰.

This paper proposes a coevolutionary optimization algorithm called DCOA.DCOA mainly focuses on how to adjust sub-population size self-adaptively so as to improvethe optimizing performance. To achieve this, a strategy is introduced which consistsof three rules: internal competition, external competition and spontaneous growthrules. These rules can control individual reproduction and elimination speed in each subpopulation.Furthermore, the adjustment can be proven globally asymptotically stable. Inthe experiments, we compare the performances of DCOA, macroevolutionary algorithm(MA) [13] and simple genetic algorithm (SGA) with typical test functions. The resultsshow that DCOA is able to find the global optimum on most difficult functions, nothingless than MA which uses simulated annealing technique. At the same time, DCOA convergesquickly, similar to SGA and faster than MA.

An immune optimization algorithm is proposed in this paper based on the immune negative selection. The algorithm NSIOA is motivatedby the negative selection mechanism in biological immune recognition. Different from the existing immune optimization methods,NSIOA constantly removes the worst solutions to get the optimal solution. Considering that removal of poor members of a populationmight lead to the loss of design information that may actually help identify better solutions in the search space, the proposed NSIOA isdesigned to keep the diversity of antibodies while removing poor members, therefore the algorithm will converge to global optimalsolution with high probability. The convergence property and the complexity of the algorithm have also been analyzed. To illustratethe efficiency of the algorithm is used in solving the travel salesman problem. The theoretical analysis and experimental results show thatthe algorithm is of a strong potential in solving practical problems.

How to dynamically adjust population size is one ofthe most important research topics in coevolutionary computation.This paper proposes a new strategy to dynamically adjustpopulation size during a coevolutionary process. Three factors areconsidered in the strategy to describe the influence on populationsize. They are the natural growth of sub-populations, the internalcompetition of individuals from same sub-population and theinteraction between individuals from different sub-populations. Itis proven that the proposed strategy is globally asymptoticallystable and also the convergence of a coevolutionary algorithm withthe strategy is investigated. Finally, the behavior of the proposedstrategy in practical use is illustrated.

Classification is an essential technology inPedestrian Detection System (PDS). Until now, single-classifierand basic cascaded classifier had been widely used in PDS;however, most of them can hardly satisfy the 3 requirements atthe same time: high detection speed, high detection rate and lowfalse positive rate. In this paper, we proposed an optimizedhierarchical classifier which can satisfy the 3 requirements. Theproposed method adopted Corse-to-fine and Early-rejectionprinciples to achieve global high performance. It consists of twohierarchies, the first one is used to quickly reject non-pedestrianobjects and select out only a few candidates; the second onemakes further verification to these candidates. Furthermore,each hierarchy was optimized with statistical models basing onexperiments; and each hierarchy is a treelike classifier which hasspecific optimization demands. At last, an overall performanceevaluation standard is proposed, and the experimental resultsshowed that the proposed classifier had better overallperformance.