Object Detection Based On Deep Learning In Driving Video

With the development of computer vision technology,especially the development of deep learning technology,autopilot technology has become more and more sophisticated.Advanced Driver Assistance System is a key step towards autopilot.Distance estimation is an important research issue in assisted driving technology.Although the accuracy of the radar measurement method of the vehicle radar is high,the cost is high.The vehicle distance estimation based on the vehicle camera is one of the cheapest alternative technologies.So,the detection estimation based on vehicle camera is researched in this paper.Based on the SSD(Single Shot Detector)object detection,a joint metric learning SSD(MLSSD for short)is proposed algorithm for object detection in driving video.The object is detected and the vehicle distance is estimated based on the marked point.In order to improve the efficiency of detection,a multi-target tracking algorithm is proposed and combined with object detection algorithm.The object detection calculation amount is reduced,the accuracy rate is improve.Finally we obtain the distance and offset angle of the target vehicle with the host vehicle in the driving video,which can help the driver to make decisions.The main work and innovations of this article are as follows.1)A distorted vehicle detection dataset is establish.The vehicle rearview camera in our research is a fisheye camera.Different with the general camera,the video frame of the vehicle captured by the fisheye camera are distorted.The model trained in the public database has a low accuracy when it used to detect the distortion data.Therefore,we annotate and build a vehicle detection dataset based on the distortion data captured by the fisheye camera.The dataset contains 5000 images,of which 4000 are training images,500 are validation images,and 500 are test images.2)MLSSD algorithm joint metric learning is proposed to detect vehicles in video.To obtain more robust features,combined with the losses of metric learning and classification,we propose a network structure based on multi-loss function that boosts the final vehicle detection result accurate.3)A combination of multi-target tracking algorithm is proposed to improve object detection performance.Based on the detection results of MLSSD algorithm,the object Detection and Tracking(DAT)framework is proposed to detect and track.Combined with a multi-objective tracking algorithm SCEA(Structural Constraint Event Aggregation),further improving the detection accuracy and real-time performance,and reducing the GPU resource consumption.Based on the camera calibration data,preliminary vehicle distance estimation is realized.The experimental results demonstrate that the proposed algorithm can detect the target in video in real time under the distorted camera and improve the accuracy.For the vehicle distance estimation,the average error is within 50cm in the test data set.