Multi-target Detection Of ADAS System Based On Video In Road Scene

Target detection plays an important role in the Advanced Driver Assistance Systems of the vehicle.When the car is driven,the driver is disturbed by the external environment,and the important targets in front of the car cannot be found in time,resulting in traffic violations or traffic accidents.Based on the urban road scene,the study is about multitarget detection technology applicable to the mobile terminal.The system needs to realize the detection of important traffic information and reminds the driver to reduce road traffic accidents and improve the safety of driving.Target detection in road scenes involves many issues such as real-time,dynamic,multi-scene,multi-target,and multi-interference factors.The video-based target detection is easy to realize the scene reappearance,which is helpful for the debugging of the model and the comparison of the test results.The traditional SSD detection model has many calculation parameters and cannot utilize the information timing association between consecutive frames of video,resulting in long detection time,low target confidence of detection,and high requirements on mobile device hardware.Hybrid model was designed on the combination of LSTM and SSD.The parameters of the hybrid model are trained on the dataset.Raspberry Pi system was written into the training parameters.When the system detects the target information,the driver is reminded by sound.Finally,realize the realtime multi-target detection system design on the mobile terminal in the road scene.The specific research content of this article is as follows:(1)This paper analyzes the video-based multi-target detection system and clarifies the scenarios and functional requirements of the detection algorithm.On the other hand,this paper comprehensively analyzes the existing target detection algorithms,and deeply analyzes the framework structure,algorithm principle and model training process of the SSD model.In addition,the data set of the training test is supplemented and data enhanced to improve the robustness of the model detection.(2)According to the actual application scenario of the detection system,this paper improves the detection algorithm.The traditional SSD model has too large a calculation parameter and a large amount of calculation.This paper adopts the framework of the lightweight network model Mobile Net-SSD,and designs a combined network LSTM-SSD for road video multi-target detection.The temporal dimension features between the video frames are extracted by the combined network,so the results of the detection model are corrected.Compared with the two detection network models of VGG-SSD\Mobile NetSSD,the average accuracy of the detection results is increased by 2%?8%,which indicates that the LSTM-SSD combined detection network model can achieve good detection results when dealing with the interference situations such as fuzziness\occlusion\etc.(3)In order to realize video target detection in a mobile device,it is necessary to construct a detection environment for the Raspberry Pi platform,and transplant the algorithm training model to the mobile platform.The system mainly includes collecting data through the camera,identifying the category and coordinate information in the image through the target detection model,and prompting the driver through the voice.After experimental testing,the combined algorithm model is transplanted to the Raspberry Pi system,which can complete the detection of target objects such as vehicles,pedestrians and traffic signs,and realize the classification and timely prompting of target objects.This research is based on the Tensor Flow development platform,and the SSD target detection,and the algorithm model was improved to form the LSTM-SSD hybrid algorithm.By comparing the effects of different models,the hybrid algorithm method effectively improves the confidence of detection and reduces the instability problem in single image detection.Then the model was written into the Raspberry Pi platform to debug and verify the whole system.When information(vehicles,pedestrians,traffic signs,etc.) is detected in real time,voice prompts will be provided.The construction of the model can provide basis and reference for real-time target detection by driverless vehicles.