S32V APEXCV 功能总结

流程	Initialize(...)，参数视功能而定	apexcv::Abs::Initialize ( vsdk::SUMat & aSrc, vsdk::SUMat & aDst )
	SelectApexCore ( int aApexId )	在初始化后调用
		apexcv::ApexcvHostBaseClass::SelectApexCore ( int aApexId )
	Process()，派生	每一帧都会执行
		apexcv::ApexcvHostBaseClass::Process()
	ReconnectIO(...)	在不重新初始化的情况下改变输入输出图像
		apexcv::Abs::ReconnectIO ( vsdk::SUMat & aSrc, vsdk::SUMat & aDst )
baselib	略
prolib	Oriented Fast and Rotated BRIEF	Orb is basically a feature matching algorithm that is very similar to SIFT/SURF but much more simple to compute. Descriptors are in a binary form and an Hamming distance can be used to find the differences between two descriptors. The output length of a descriptor can be 16, 32 or 64 Bytes. The added value of this implementation is the fact that the pattern used to generate the descriptors needs to be given by the user.
	Canny Edge Detector	It calculates the X and Y gradients using 3x3 Sobel filters, computes the magnitudes, and performs non-maxima suppression, and double thresholding to determine good edge points.
	GFTT/HARRIS Corner Detector	The algorithm is a fixed point implementation of Harris Corner [4] and GFTT corner [7] and is similar to the function goodFeaturesToTrack in OpenCV.
		1. Computes a 16-bit corner response image from an 8-bit grayscale image using Harris Score Formula [4] or Minimum Eigen Value (GFTT) [7]. 2. Performs non-maxima supression in a 3x3 neighborhood. 3. Removes any corners that are below threshold. GFTT Threshold = QualityLevel ? MaxEigenValue. Harris Thresh- old is specified by users. 4. The remaining corners are sorted by the strength in the descending order. 5. Minimum distance filtering will be applied. A strongest corners within minimum distance radius will be kept. Other corners will be discarded.
	HOG Object Detector	The algorithm is a fixed point implementation of [1] using a simpler HOG feature. It is similar to the function HOG- Descriptor::detect in OpenCV.
	Aggregated Channel Feature Based Pedestrian Detector	The aggCF detector takes a 24 bit RGB image and calculate aggregated channel feature pyramid which is divided into octaves. Each octave includes real scale and approximation scales with different scale sizes. Each scale includes L, U, V, magnitude and histogram of gradients (HOG) as channel features. HOG includes 6 gradient bins and can be calculated using both bi-linear and tri-linear interpolation based on pre-trained detector model parameter. The structure of pyramid including number of octaves/scales are determined by pre-trained detector model. Once feature pyramid is calculated, aggCF detector perform pedestrian detection based on decision trees in pre-trained detector model, apply non maximal suppression (NMS) on bounding boxes of detected pedestrian. The detection is performed using sliding window approach, the size/stride of search window is also defined in pre-trained detector model.
	Hough Line Detector
	FAST9 corner detection	APEX-CV FAST algorithm implements FAST corner detection [6] to extract feature points from 8-bit grayscale source image. When "nms" flag is enabled, FAST corner score will be computed to perform non-maxima suppresion post processing step.
	Gaussian Image Pyramid	子主题 [http://cdn.processon.com/5d9e907de4b0893e99232186?e=1570676366&token=trhI0BY8QfVrIGn9nENop6JAc6l5nZuxhjQ62UfM:46_N_-h5EUAfUBGhZAPFSHn19hk=]
	Multi-scale Gaussian Image Pyramid	APEX-CV Pyramid Multi supports 4 scales downsampling of an source image through one APEXCV call. The source image is convolved with the Gaussian filter in the same way as image pyramid kernel then down sampled. Each APEXCV call will produce [1/2, 1/4, 1/8, 1/16] 4 scales pyramids.
	Laplacian Image Pyramid	The Laplacian Image Pyramid is generated with the help of a Gaussian Image Pyramid. For each of the pyramid levels, the laplacian image is obtained by convoluting the Gaussian Image Pyramid level with a 5x5 Gaussian kernel and subtracting it from the input image. There is also an final output image whis can be used together with the Laplacian Image Pyramid to reconstruct the original image. This image is the convolution result of the last pyramid level.
	Affine Transformation
	Image Remap	The APEX-CV Remap maps one image to another from a floating point lookup table. The algorithm is similar to the function remap in OpenCV . The current implementation of apexcv::Remap supports image sizes that are an even multiple of the number of CUs (64). So supported images are 128, 256, 384, 512, ... , 2048. The current version of remap supports bilinear interpolation only.
	Histogram Equalization	The APEX-CV Histogram Equalization transforms the pixel values of an image so that the resulting histogram uses the full range of grey values equally. The algorithm is similar to the function Histogram Equalization in OpenCV. The current implementation of apexcv::HistogramEqualization supports image sizes which have a width multiple of 4.
	Image Resize	The Image Resize performs a vertical and horizontal resize on the input image, according to the required size of the output image. Note that vertical and horizontal scale factors (and directions) are independent. It is similar to the function resize() in OpenCV. Currently, all even image dimensions are supported from 64 to 1024 for both input and output. The current version of resize supports bicubic interpolation only.
	Tone Mapping Operation	The Tone Mapping Operation generates an LDR image from an HDR image by compressing the HDR’s dynamic range. The APEX implemetation of TMO is a fixed-point implementation of tone mapping operation as described in [2]. The implementation uses integer data and fixed-point arithmetic for all calculations in TMO to reduce computational and memory cost.
	Single-Scale Lucas-Kanade Optical Flow
	Multi-Scale Lucas-Kanade Optical Flow
	LBP Face Recognition	The APEX LBP face recognition is similar to Local Binary Patterns Histograms in OpenCV. The algorithm will run APEX-LBP train process to extract an LBP descriptor for each grid cell. The process will take as input an unsigned 8 bit image and output an 8 bit descriptor for each grid cell. Then the APEX-LBP predict process will be executed to compare the test descriptors with the model descriptors in order to find the closest descriptor. This predict process will output an unsigned 16 bit value representing the closest ID and a set of signed 32 bit distances.