Night-time Preceding Vehicle Detection Based On Monocular Vision

Night time vehicle detection has always been a challenging task in the field of computer vision.External physical characteristic of vehicles are hard to notice under poor lighting conditions,which serves as a major cause for frequent night-time traffic accidents.Night-time vehicle detection error rate is high,the detection speed is slow,therefore,a method for preceding vehicle detection based on monocular vision is hereby proposed to improve safety of driving at night.This thesis mainly consists of three parts:(1)Lamp detection method which based on CenSurE features,due to manually set the threshold and filter linear sampling,resulting in feature points sparse and noisy.We obtain CenSurE responses with logarithmic scale and eliminate unstable feature points with the use of multilevel Otsu thresholding.Next,locate the lamp area in HSV color thresholding.In the end,vehicle detection is completed by lights matching according to matching constraints.It can be concluded from the experimental results that our is able to reduce noise,exhibit 8.6% improved Precision to 41.2%,12% improved Recall to 53.3%,but still detected some non-vehicle light sources.(2)In this task specific application,light sources are recognized as one of the most important features that have significant influence on model performance(Precision and Recall).Different from original Faster RCNN,which selects negative samples randomly,proposed method mining hard negative examples base on CenSurE features.Negative examples which contains non-vehicle light sources are considered as hard negative examples.These examples are employed in model training phase which expect to help Faster RCNN to distinguish vehicle light sources and non-vehicle light sources.Experiments results shows that Faster RCNN trained with hard negative examples exhibit 8.1% improved Au C(area under the P-R Curve)to 84.4%.(3)In order to improve the detection frame rate,a PCA based model compression method is proposed to compress the weights between fully connected layers.On the other hand,to guarantee model convergence,an optimized model training pipeline is introduced and employed to train the compressed model.The final optimized model is a compressed Faster RCNN trained with mined hard negative examples.Experiments results shows that,the final optimized model have around 5 million less learnable parameters and the model detection frame rate is increased from 17 fps to 24 fps.Experiment results illustrate around 9.8% improvement in Au C to 86.1%.Moreover,besides use experiments data which are similar to the training dataset,more challenging data are used to further evaluate models' robustness and generalization capability.Through empirical studies on experiments results,the optimized Faster RCNN exhibit a good degree of applicability in the context of night time vehicle detection task.