Binocular Infrared Significant Target Extraction And Its Hardware Implementation

With the development of night vision technology,infrared detectors are widely used in unmanned driving,machine navigation,virtual reality,medical image processing and other fields.A single infrared camera can only capture the surface of the image and unable to get the deep information.So,the use of binocular stereo night vision technology to achieve three-dimensional reconstruction of the image scene is the developing trend.The correction of the polar line and detection of significant objects have been studied in this paper,and the hardware platform of extraction for binocular infrared significant target has been designed and built.The binocular stereo night vision system uses two infrared cameras to acquire images.Due to engineering errors,this paper studies a method of pole calibration based on single transformation matrix to ensure the both two cameras in the same level.This method uses the left image as a reference map to linearly transform the right image,so that the transformed right image and the original left image are corrected image pairs.For saliency detection,this paper proposes a global contrast detection method based on LARK regional features.This method uses the local feature covariance matrix as the eigenvalue for significance detection.By introducing the spatial weighting information and the luminance weighting information,the local salience detection is extended to the global significance detection.Aiming at the saliency detection algorithm based on background information,this paper proposes an effective distance estimation model with utilizing the distance between the pixels to extract the boundary super-pixels and using the boundary contrast to refine the boundary clues.This method can achieve the ultimate effect of improving the detection model which is based on background information.A hardware platform based on multicore DSP6678 and FPGA is designed and built to realize the system hardware.The algorithm of single transformation matrix and saliency detection can be integrated into the hardware and the experimental results show that the system can achieve the real-time processing,and extract significant target information well.