No abstract

Presented is an intensity-based feature extraction method for pedestrian classification in far-infrared (FIR) images. The underlying idea of the method is that only intensity differences between neighbouring pixels can represent both the direction and the magnitude of the gradient, as FIR images are characterised by monotonic grey-level changes. A new intensity-based feature called the histogram of local intensity differences (HLID) is introduced which is a modified version of the well-known histograms of oriented gradients (HOGs) feature. Experiments show that the HLID is more suited to FIR images than HOGs in terms of both accuracy and computational efficiency.Introduction: The selection of a feature extraction method is the key point of a discriminative classification technique that determines an optimal decision boundary between positive and negative classes through sample training. Since the feature of histograms of oriented gradients (HOGs) was introduced by Dalal and Triggs [1], several studies have shown that a HOG outperforms all other techniques of single feature extraction for pedestrian classification in visible spectrum images [2]. Currently, the HOG is widely used for pedestrian classification in not only visible spectrum images but also far-infrared (FIR) images [3][4][5]. However, it was developed on the basis of visible spectrum images. It should be noted that the characteristics of FIR images differ from those of visible spectrum images. Temperature changes within the same object are not significant, and FIR is hardly affected by the texture of the object surface; therefore the FIR camera captures low-frequency images, and as a result, it generates images characterised by monotonic grey-level changes. These attributes are well illustrated in Fig. 1. As shown in the mesh plot, the local intensity variation of an FIR image is smaller than that of a visible spectrum image.

No abstract