Small And Weak Target Detection And Intelligent Analysis

With the rapidly development of digital image processing, computer visual and pattern recognition, target detection, tracking, and recognition technology has been widely used in our real life.All detection systems are designed to provide a robust detection and recognition method, both influenced by weather such as fog or small static target, or pedestrian, vehicles and other common moving target, we hope to have a lower false detection rate and better recognition rate. To achieve this goal, the system application environment should be taken into full consideration, thus we can provide excellent performance of specific detection and recognition algorithm.In this thesis, we will introduce a real-time detection system that can detect both static and moving weak and small target, the system also provide intelligent analysis. The system consists of three modules, preprocessing, target detection and intelligent analysis. The preprocessing module is mainly completed image enhancement before detection, it aims at improving the SNR of images such that the system can better deal with the situation where the visibility and contrast are low, and it also can improve the detection rate of small and weak target. In this thesis, we use image enhancement algorithm based on the theory of Retinex, and make use of CUDA parallel programming model, the method is realized on the latest NVIDA embedded GPU platform Jetson T K1 to increase the speed of preprocessing. In the next step, target detection, we propose a kind of dual channel detection model and illumination correction algorithm, and the real-time static and moving target detection is realized in the TI company’s DM8127 image processing platform. In the end, DM8127 platform will encode these video frame sequences based on H264, then sends the encoded video stream and target detection results to the backend host, the host will analyze these targets based on HOG and SVM algorithm, to identify whether the targets are pedestrians. Traditional pedestrian recognition algorithm has to do multiple scales and intensive scanning to an entire image, but in our method, we only extract the HOG feature for the areas which include targets. In this way, we can reduce the complexity of the algorithm while not affect the recognition effect, and make sure the system is real-time.Through testing this real-time weak and small target detection and intelligent analysis system, it shows that the system has good robustness, and can overcome the influence of weather factors, for example, the foggy day, and illumination conditions. Whether it is a weak and small static target or a moving target, the system keeps good performance, especially for the height of 1 cm or above, more than 3 cm in length static target, which has different color with background. Its missrate can reach less than 5% within 50 meters, and the average time interval between false positives targets is more than 2 hours. For 1080 p input video, the whole system’s processing frame rate is about 20 fps, the real time detection and recognition can guarantee the system applied into the real life.