Research On Intelligent Moving Chassis Following Characteristics Based On Visual Recognition Technology

With the rapid development of artificial intelligence technology,intelligent chassis plays an increasingly important role in areas such as sanitation,medical treatment and fire fighting.A lot of work requires self-following chassis to carry different equipment to cooperate with the leader in front.Traditional self-following technology mainly relies on ultrasonic positioning,ultra-bandwidth radio positioning technology,etc.However,the ultrasonic positioning distance is limited and the following ability is not enough in complex environment.Ultra-bandwidth radio technology requires the front follower to wear a device to complete the following.This paper aims to solve the problems that traditional self-following technology has poor anti-interference ability,small application range,or the leader needs to wear extra equipment.The thesis carried out the following work:1)Completed the establishment of intelligent mobile chassis identification scheme.Firstly,the tracked intelligent chassis is selected as the mobile platform,YOLO-V4-TINY is adopted as the self-following recognition scheme,and the whole body of the human is used as the recognition object.Open CV is used as gesture image processing scheme to complete gesture recognition.Finally,the whole development environment and the program running environment were built.2)Train and optimize the detection model.Firstly,the pedestrian detection model is trained and tested.The pedestrian image collection and data set establishment were carried out,and the convolutional neural network was used to complete the training of the detection model.TensorRT was used to optimize the model,and the test was carried out on the data set and the evaluation criteria were used for analysis,and the real-life test and verification was carried out.The detection accuracy of the proposed detection model exceeds the original YOLO-V4-TINY detection model by 12.7%,but the detection speed is reduced by 2.9%.After acceleration through TensorRT optimization,the detection speed is improved by 10.4%and 40%compared with YOLO-V4-TINY.Secondly,gesture recognition technology based on Open CV is used to complete gesture recognition and classification.The palm image processing and hand feature extraction are realized.The data set of gesture image is established,and the processed data set is classified by support vector machine.The recognition accuracy of no class reached 98%,and the average accuracy reached 98.9%.The identification and verification under various light in the actual scene is completed.3)Completed the development of the target recognition and tracking strategy in multi-pedestrian scenarios.Based on the Deep Sort tracking algorithm,the function of selecting the specified target in the case of multiple pedestrians in front of intelligent chassis is developed,and the method is tested and verified on MOT16 data set.In this data set,the MOTA,MOTP and FPS reach 63.2%,72.1%and 26.5 respectively.The function of selecting the specified target is completed.4)A real car test was carried out.The following mode of intelligent chassis is studied,and the following free test of intelligent chassis to the target in front and the following test of specific channel are completed in the actual scene.The final relative errors of following are all within 10%,which verifies the stability of intelligent chassis following effect.