Research On Multi-Object Detection Algorithm Based On Vison In Unmanned Vehicle Environment

The key for vision-based multi-object detection on unmanned vehicle environment is the ability to accurately determine road image pixels as the multi-object region or not by designing model.Vision-based multi-object detection is very challenging for heavy traffic flow,big data and continuously changing background.Deep learning of vision-based cannot require designing different features for specific multi-object on application of multi-object detection and act in accordance with the process of human brain processing visual information,making it closer to artificial intelligence,thus multi-object detection algorithm based on deep learning of vision has become a research hotspot.In this paper,for the specific application of unmanned vehicle,we study how to detect multi-object based on deep learning of vision in noisy road environment and the proposed algorithm for multi-object detection provides a high accuracy and robustness.The main work and research results are as follows:In order to improve the processing speed of multi-object detection and reduce the interference of irrelevant information,according to the driving parameters of the unmanned vehicle,getting the region of interest(ROI)from images of the vehicle-mounted calibrated camera captured,an algorithm of multi-object regional extraction combining segmentation and saliency of visual attention mechanism is put forward,which is based on region instead of based on sliding-window and meet the requirement of real-time for multi-object detection of unmanned vehicle.First,the image is divided into different and overlay regions from global view.Then we measure the saliency value of regions to get object region proposals from local view.For saliency,we propose an innovative cue to calculate the characteristic of the image.In order to accelerate the processing speed of the video,we also combine with the traffic parameters to estimate the position of object windows.We do experiments on PASCAL VOC dataset and real datasets collected on city streets.Compared with state-of-the-art object region extraction methods,the proposed method generates a very small set of category-independent object windows less than 30 and shows better results on speed and accuracy.In order to make the feature representation of multi-object much closer to human intelligence and imitate the process of human brain dealing with information,this paper proposes a method for multi-object detection by combining human brain multi-mechanism.The multi-mechanism including saliency,sparseness,locality and depth,is used to represent object-level feature.The algorithm first uses non-negative sparse coding locality-constraint linear coding to obtain sparse and local feature,and then uses salient pooling and local grouping to obtain abstract features expression,repeating this process like neural network process to get the deep feature,which is benefit for feature representation.For the last output layer,it uses Gaussian pyramid saliency pooling to ignore the size of input.Finally,features are used with classifier to detect multiple objects.Compared with state-of-the-art object detection methods,the proposed method shows better performance.Implement object regional extraction algorithm via segmentation and saliency of visual attention mechanism and multi-object detection algorithm by combining saliency,sparseness,locality and depth.Experimental results demonstrate that it is robust to lighting variations and heavy traffic.It provides the better performance when compared with several state-of-the-art methods.