Intelligent hough transform for object detection and tracking

Abstract

In this thesis, different approaches for enhancing the Hough-based technique for shape detection have been investigated. Most techniques have been successfully realized in a DSP platform and ARM9 platform for two industrial projects: Advanced Driver Assistance System (ADAS) and Watch latency detection system. The major part of this thesis is based on the Hough Transform technique, which is a method of detecting characteristic contours or shapes by exploiting mutual constraints between parameters and points lying on the target contour. Due to high memory usage and consumption of computing power, normally the Hough Transform is not suitable in real time and embedded system applications. Besides, the noise from camera and obstructions of real sense will cause miss or false detection results of any Hough detection system. In the literature, a number of researchers proposed to optimize the use of hardware for Hough transform such as to run the computation of accumulator array in parallel or store data in an optimized structure of memory. Or some of them proposed to use feature based method to de-noise the image by analyzing local features, or use a model to conduct fitting to reduce the possible feature points before adopting them to the Hough transform. In our work, we have proposed a block based method to eliminate the non-lane block using their localized features, such as mean and standard deviation from the mixed colour histogram. By collecting the features from numerous samples of Lane and non-Lane blocks, we form a specified filter. Besides, we have proposed an inverse Hough model to conduct curve fitting for tracking the lane after detection. This provides a fast camera based method to identify the region of interest from the captured screen and also provides a defined area for tracking status that increases the detection accuracy to 90% and reduces the hardware requirement by using only 500 MHz core. For the watch latency detection system, we have suggested to detect the existence of watch by detecting the pattern on the holder. To reject the noise for background moving parts on the watch, we made use of an average background method to build up a codebook for each pixel. Then using the frame difference over time and the nature of second hand movement, the hand position of the watch can be detected. This also facilitates the estimation of the latency of the watch accuracy up to 1 second. Moreover, the General Hough Transform is optimized and adopted to recognize a selected pattern of the holder under embedded system environment with less than 1 fps.