Research And Implementation Of DSP-based General Object Recognition

In the domain of intelligent video processing, object recognition is fundamental to understand the semantic information of videos. In the current market, many domains such as automation, intelligent assistance, robots, automated driving, has been more and more dependent on object recognition technology. However object recognition algorithm is a kind of high degree of intelligence, so it is very complex. With categories increasing, it is difficult for object recognition algorithm to find a common classification model.General object recognition research the common characteristics of each object and the mainstream is to use rectangles to mark the location of each object in the image. Many general object recognition algorithms explore a variety of different viewpoints of commonalities between individuals. They partially solve the problem that the classical object recognition algorithms can not quickly identify many types of objects. So they promote the practical application of object recognition algorithm. Because differences between different objects is very large and there are a wide range of objects, general object recognition algorithm still has a lot of shortcomings and deficiencies. Existing general object recognition algorithms are based on training methods, training a large number of parameters are not suitable for embedded applications; by the environmental impact, it is difficult to adapt to complex scenarios; with complex computation, it is difficult to meet the real-time requirements of embedded system. This article start from the practical application, focusing on how to efficiently assess the general target, has mainly done the following work:1. To overcome the problems and shortcomings of existing algorithms, this paper presents a new contour edge evaluation method. In this method, using the general characteristics of common convexity between objects as our breakthrough point, we build an unsupervised method. This method does not require the training set, so it is adaptable. Based on the characteristics of the convexity, it is very fast. In aspect of real time, the new algorithm has greatly improved than previous methods.2. Existing candidate box evaluation algorithm is slow, this paper proposes a new method based on dynamic programming. The proposed method provides a better balance between the accuracy and speed.3. In order to transplant the above-mentioned contour edge evaluation method and candidate box evaluation algorithm to DSP, Using TI's Sys Link assembly, this article set up a general object recognition system based on the DM8168 platform. In the system, the inter-core communication between individual cores is based on messages, data exchange between the individual cores is based on shared memory, each thread runs in parallel through multi-buffer queues.4. Taking special structure of C674 x DSP kernel into consideration we transplant algorithm to the DSP. In this paper, the algorithm of each module has been redesigned and optimized to reduce memory consumption and running time. The overall performance of the system has been greatly improved after optimization.