Vehicle Detection And State Judgement At Nighttime

Nighttime vehicle detection and state judgement is based on front-mountedcamera, it means detecting vehicles in front and making decision of their stateaccording to their taillights’ state. The research of this paper mainly contains two parts:Firstly, a new method is proposed related to the key problem of nighttime vehicledetection, which is taillight detection; secondly, this paper presents severalimprovements about other steps of this problem.The existing solution to the problem of nighttime taillight detection relies toomuch on the input video’s circumstances, which can only detect a few kinds oftaillights with specific types, and the detecting threshold varies depending on thedistance. To solve these problems, this paper proposes an improved algorithm. Firstly,the algorithm uses ROI to define the detecting area; secondly, use thresholdscalculated from a large number of sampling data with wide ranges to filter the image;at last, pick up the white zone to get a mount of taillight hypothesis and dopreliminary verifications. In the step of hypothesis verification, the algorithm firstexpands the inner rect to get the approximate extent, and then expands the whitezones layer by layer to make coverage of the halo zone precisely, and verifies the halozone creatively based on a fact that “the pixel farther from the white zone border haslower luminance”, and finally gets the taillight detection result. The experiments showthat the algorithm of this paper has a better adaptability on taillight’s colors andshapes with a higher detection rate compared with the existing algorithm.In the matter of nighttime vehicle detection and state judgement, existing taillighttracking algorithm lacks a proper way of dealing with the missed tracking case, andhaven’t considered the taillight state changing situation; in taillight pairing process,the existing algorithm appears to be too strict, requiring the object directly facing thehost car which leads to narrow detectable angle; in the step of state judgement, theexisting algorithm does not take full advantage of taillights history information andthe characteristics of the taillights in special cases. To solve these problems, this paperproposes an algorithm. Firstly, in the step of tracking, the algorithm uses two queuesto track taillights and filter the unwanted objects, and uses least square method tomaintain the stable taillight objects and the consistency during the detection process; in the step of pairing, this paper proposes a method which uses a series of loose rulesto pair the taillights and only does the pairing process once so that the algorithm candeal with the taillight light signal changing cases (such as turning signal), at this point,the algorithm gets the result of the vehicle detection; in the step of state judgement,this paper uses both the luminance feature and the color feature, and judges the stateof the front vehicle by the accumulating the change of these two features during thelatest3frames, the result proves to be good. The experiments show that the algorithmproposed by this paper is able to achieve the goal, the vehicle detection algorithm hasbetter detection rate and adaptability compared with the existing algorithm. The statejudgement algorithm is accurate. The whole algorithm is efficient and real-time.