Transportation Mode Detection Based On Deep Convolutional Neural Network

Accompanied by the growth of urban population and the acceleration of urbanization,problems of transportation and daily commuting have gradually become an obstacle to the city's further development.Scientific traffic planning and management plan is an effective way to solve the increasingly complex traffic problems,and the basis of formulating it is a large amount of daily trip information obtained by residents' trip surveys.As an important content of the trip survey,transportation modes not only reflect urban traffic problems,but also embody the residents' travel characteristics and activity patterns.Therefore,it has been a trend to obtain transportation modes more efficiently and accurately.Because the traditional trip survey methods suffered from many problems such as lack of objectivity,high cost and long implementation period,the shortcoming of it becomes increasingly unbearable with the development of urban transport.In recent years,the popularization of GPS equipment and services as well as the rapid rise of the intelligent industry have jointly promoted the researches on automatically collect and recognize the transportation modes.Benefits from the high quality,wide coverage and fast update frequency of GPS data,GPS trajectories detection have become a novel way to solve the problems from traditional trip survey,and extract transportation modes from GPS data turn into a difficulty and focus both in research and implementation.The current research mainly has difficulties in feature selection,timing analysis and transform point recognition,which impede the improvement of the accuracy of transportation mode recognition.Based on the above difficulties,this paper presents a deep convolution network scheme for detecting transportation mode from GPS trajectories,which including walking,bicycle,bus and car.The main contributions are shown as follow.Firstly,deep convolution network is designed to solve the problems of feature selection.By using the ability of deep convolution network to learn data characteristics autonomously,the detecting algorithm can avoid feature selection as well as judging its validity.All data needed for scheme only from GPS trajectory data,rather than using auxiliary data such as road network or other sensors,which cause a better scalability for the scheme.The proposed algorithm is verified by GPS data from Geo Life project,and results indicate that the deep convolutional network can effectively detect four transportation modes from mode segments,in which walking owns the highest accuracy and recall,meanwhile cars and buses both have 85% detect accuracy.Finally,the detection results of varies trajectories at different duration,distance,sampling rate and moment are compared,which further illuminate the validity of the algorithm.Secondly,transportation mode detection method for mode segments is propose based on deep convolution network and local moving window.The algorithm first extract the timing characteristics from each window segment,then these features are identified by network,next a post-processing method is designed to advance the accuracy of detection,finally this result is compared with those without using timing features,which demonstrate that timing features can effectively enhance the recognition ability of buses and cars.Thirdly,online transportation mode detection method for multi-mode trajectories is constructed to avoid the difficulty of transform point recognition.Based on the algorithm presented above,this scheme improve the post-processing method ground on the factual rules,and is able to make a result of detection before trajectory is completed.Meanwhile,the moving window based feature extraction method limit the influence of transform points to a few segments,which satisfies the prerequisite for online mode detection for multi-mode trajectories.Experimental results indicate that F-score of all modes detection results are above 80%,of which the F-score of walking and car reach 90%,verifying that this method is able to detect online multi-mode trajectories effectively and efficiently.