The existing deep learning based infrared and visible image fusion methods have too many parameters and require large amounts of computing resources and memory. These methods cannot meet the deployment demand of resource constrained edge devices such as cell phones and embedded devices. In order to address these problems, a light-weight image fusion method based on SqueezeNet was proposed. SqueezeNet was used to extract image features, then the weight map was obtained by these features, and the weighted fusion was performed, finally the fused image was generated. By comparing with the ResNet50 method, it is found that the proposed method compresses the model size and network parameter amount to 1/21 and 1/204 respectively, and improves the running speed to 5 times while maintaining the quality of fused images. The experimental results demonstrate that the proposed method has better fusion effect compared to existing traditional methods as well as reduces the size of fusion model and accelerates the fusion speed.

Movidius neural computing stick is a USB-based deep learning inference tool and a stand-alone artificial intelligence accelerator that provides dedicated deep neural network acceleration for a wide range of mobile and embedded vision devices. For the embedded application of deep learning, a near real-time pedestrian target detection method based on Movidius neural computing stick was realized. Firstly, the model size and calculation were adapted to the requirements of the embedded device by improving the RefineDet target detection network structure. Then, the model was retrained on the pedestrian detection dataset and deployed on the Raspberry Pi equipped with Movidius neural computing stick. Finally, the model was tested in the actual environment, and the algorithm achieved an average processing speed of 4 frames per second. Experimental results show that based on Movidius neural computing stick, the near real-time pedestrian detection task can be completed on the Raspberry Pi with limited computing resources.

Feature representation and metric learning are fundamental problems in person re-identification. In the feature representation, the existing methods cannot describe the pedestrian well for massive variations in viewpoint. In order to solve this problem, the Color Name (CN) feature was combined with the color and texture features. To extract histograms for image features, the image was divided into zones and blocks. In the metric learning, the traditional kernel Local Fisher Discriminant Analysis (kLFDA) method mapped all query images into the same feature space, which disregards the importance of different regions of the query image. For this reason, the features were grouped by region based on the kLFDA, and the importance of different regions of the image was described by the method of Query-Adaptive Late Fusion (QALF). Experimental results on the VIPeR and iLIDS datasets show that the extracted features are superior to the original feature; meanwhile, the improved method of metric learning can effectively increase the accuracy of person re-identification.