Research On Pedestrian Detection And Tracking Method Based On Automotive Vision

With the increase of car ownership,the safety issue of pedestrians is seriously brought to the forefront of public attention.Pedestrian detection and tracking system based on on-board camera is a part of pedestrian protection system to recognize environment,which holds a great significance for reducing the frequency of accidents and alleviating the extent of injury.Traditional studies tend to treat pedestrian detection and tracking as two separate parts,thus ignoring their relation.Based on pedestrian detection algorithm and pedestrian tracking algorithm,this essay studies the key technologies about pedestrian detection and tracking system.The essay firstly presented the aggregated channel features module and then improved this detection algorithm.The theories of the aggregated channel features and fast feature pyramids are analyzed to put forward a multi-scale detection method combining feature pyramids and multi-scale detectors.The regional search strategy is proposed to improve detection efficiency on basis of distribution of targets in image.A fusion algorithm of staged classification for weak classifiers was made to reduce classification workload based on the classification principle o f detector.The theoretical knowledge of the continuous convolution operator tracker was analyzed in detail,and a comprehensive evaluation index was set up by considering the response characteristics of filters to adjust the update strategy to prevent the problem of template drift in object tracking.The Autoregressive Integrated Moving Average time series model(ARIMA)was introduced to fit the tendency of location and scale of targets,then a pedestrian prediction model was proposed to re-track the target under a long-term occlusion.The essay put forward a program of pedestrian detection and tracking system to realize multi-objective detection and tracking with specific parameters.With MATLAB and Visual Studio C++ joint-programming,the multi-scale detectors were trained and our detection algorithm was evaluated based on the Caltech dataset s.The results indicate that the improved algorithm can effectively reduce the false positive windows and increase the detection speed.Compared with other tracking algorithm,the proposed tracking algorithm showed more excellence.The ARIMA model was proved to be comparatively accurate compared with the ground truth from the Caltech datasets.At last,the essay conducted experiments on the Caltech datasets,which results show that the presented pedestrian detection and tracking system can accomplish the multi-scale and multi-object detection and tracking task to get quite reasonable results about the location and scale of targets.