Design And Implementation Of Pedestrian Detection System Based On Multi-feature Fusion

Intelligent video analysis has been widely applied to the practical projects, such as “Safe City”, “Smart City”, etc., because it can ensure the safety of the people's lives and property and it is an effective method for intelligent traffic management. As the research focus of intelligent video analysis, due to low detection rate and low accuracy, pedestrian detection cannot effectively work in practical monitoring environment. Two factors can influence the effectiveness of pedestrian detection. First, single pedestrian feature cannot accurately present abundant body posture and exclude the complexity and similarity of background interference. Second, pedestrian feature dimension is high and the computational complexity of sliding window scanning method is high.To improve the accuracy of pedestrian detection and further improve the processing speed of video monitoring in actual monitoring environment, the thesis carried out the following studies:(1) Using multi-feature fusion to extract pedestrian feature. By fusing the pedestrian gradient histogram feature-HOG(Histogram of Oriented Gradient) and based on histogram shearlet transform-HSC(Histogram of Shearlet Coeffcients) feature, we can enhance the pedestrian feature representation and improve the classification ability for feature classifier.(2) Multiple rounds training on false detection samples. During the training process, by adopting SVM to train pedestrian features, focusing on the error detection samples on the previous round, and adding the samples to the sample set for further training, the misjudgment rate of classifier can be significantly decreased.(3) By using Gaussian model and two frames difference method to build a background model, we can extract the interest regions, reduce the error rate of pedestrian detection caused by complex background, and improve the pedestrian detection processing speed by reducing the size of the detection area.(4) Proposing the target tracking algorithm to record the trajectory of pedestrian and increase the processing speed and robustness of pedestrian detection system. For each detected pedestrian, we use target tracking to detect pedestrian's position in the following video sequence.According to the above technology routes, we has implemented the pedestrian detection system based on the multi-feature fusion, multiple rounds training, background model construction, and the target tracking algorithm. Finally, to evaluate the performance of our system, we conduct extensive experiments. Evaluation results demonstrate that: The design scheme we proposed optimizes the accuracy and speed of pedestrian detection to a certain degree extent and has a positive significance for the research of pedestrian detection.