Research On Moving Human Targets Detection And Tracking In Traffic Surveillance

Human Moving target detection and tracking is the current hotspot in image processing and computer vision research field, and it is widely used in security monitoring, customs anti-smuggling, anti-terrorism, military and some other fields. In order to achieve the function of detection and tracking the work usually be divided into three steps: (1) moving target detection; (2) human target recognition; (3) target tracking. This paper concentrates on the application of traffic surveillance and starts from the three major steps, analysis the algorithms of each functional part respectively and improves the algorithm according to the shortcomings of the current proposed algorithm, ultimately builds a complete moving human target detection and tracking system with software and hardware platform. The main contributions are:1. During the Gaussian kernel density based moving target detection, the current model can not distinguish between moving objects in the background and periodic interference, which could lead to incorrect model, we put forward a eight neighborhood based model by analyzing the different reasons for periodic disturbance and movement, build the target pixel and its eight neighboring pixels together, to distinguish the moving objects and the periodic interference exist in the background by determining the differential relationship of gray value for the current frame and background pixels. The experiment results show that it is an effective method for distinguishing the moving objects and the periodic interference and can provide a good background sample for Gaussian kernel density based moving target detection.2. In the research of human object recognition algorithm, put forward a classification error iterations based threshold detection method (TDEI algorithm) according to the excessive training sample problem for traditional Adaboost algorithm. In the calculation of the classification error of each round ,the training method firstly order the positive and negative eigenvalues samples, and then exclude the edge features within the fault-tolerant according to the Adaboost training error rate of last round, lastly search the error threshold in the remaining edge interval eigenvalues. Compared to traditional Adaboost training method as a full search method, TDEI algorithm can reduce the search range of the classified error threshold up to 80% or more, it can greatly save the time required in training sample.3. Present a goal re-acquisition algorithm based on differential information according to the problem of goal missing when the target is partly or completely blocked in target tracking by using Mean Shift algorithm. The method firstly using linear prediction mechanism to predict the possible target area and then use the image differential information to determine the candidates , ultimately determine the tracking target by Bhattacharyya coefficients.4. In the image pre-process of establishment for system hardware platform a FPGA-based median filter method is proposed to read data based on the principle of FPGA, will cache the data obtained during the comparison through the D flip-flop, and eventually reduces the comparison number of 3 * 3 median filter to 13 times, greatly saves the resources of FPGA hardware and improves the algorithm efficiency.