A Review on Video Based Vehicle Detection, Recognition and Tracking

On Road Vehicle Breakdown Assistance Finder Project

 A Review on Video Based Vehicle Detection, Recognition and Tracking

Motion Based Models: Motion based methods extract moving vehicles based on motion from background. Motion based methods includes Temporal frame differencing and Back ground subtraction.A. Frame differencing:Frame differencing is least complex and quickest method. Pixel wise difference is figured between two back to back frames. Moving foreground regions are determined using a threshold value [4]. The detection can be improved using three consecutive frames. Dual inter frame subtraction followed bitwise AND is performed to extract the moving object [3]. B. Background subtraction: Foreground objects are separated by computing pixel wise distinction between the present image and the static background image [7]. The information about the background is accumulated to produce the background model. Background can be parametric or non-parametric. •Parametric Models: Parametric model uses a uni– model probability on each pixel and update the distribution parameters. Frame averaging: Frame averaging is a conventional averaging technique where a set of frames are averaged. The resulting background model will be subtracted from consecutive frames this technique has high computational efficiency but has tail effects [13]. The accuracy depends on N which comes at the cost of memory requirements. Single Gaussian: Background is modeled recursively using single Gaussian. This improves robustness and reduces the memory requirements. The background is computed recursively in terms of cumulative running average and standard deviation [5]. Each pixel is distinguished into background orforeground depending on the pixel position in Gaussian distribution. This model reduces the cost but tail effect still persists. Median Filter: The background is evaluated by locating the median value for each pixel from a set of frames [13]. This technique is adopted when background pixels will not vary rapidly with time. A recursive approximation approach estimate the median using recursive filter which increases or decreases by one whenever input pixel has a value greater or less than estimate and is not changed if equal. Gaussian Mixture Model: Gaussian mixture models each pixel as a blend of two or more Gaussian temporally updated [21]. The stability is evaluated to estimate the distribution as table background or short- term foreground process.

https://codeshoppy.com/shop/product/on-road-vehicle-breakdown-assistance-app/

 •Non parametric Models: Non parametric background models use pixel history to construct a probabilistic model of the observation using recent samples of pixel values and do not consider pixel value as particular distribution [13]. Kernel Density Estimation (KDE) and codebook are non-parametric models[16]. Kernel Density Estimation (KDE): the nonparametric KDE characterize a multimodal probability density function [14]. The probability of each background pixel is estimated from recent samples using Parzen window. Codebook Model: Parameters represented by probabilistic function is replaced by a set of dynamically handled code word to model background [15]. Quantization or clustering is applied after this. Each codebook contains a number of code words. The new pixel is classified as background if the value of the pixel belongs to the code word range else classified asforeground. 
Vehicle Tracking Vehicle tracking is employed to obtain the vehicle trajectory by identifying motion dynamic attributes and characteristics to predict its position in subsequent frames [1]. Vehicle tracking can be categorized into three models: motion based tracking, region based and feature based tracking. Model Based Tracking: Model based tracking uses prior knowledge to createa geometric model of vehicle, which can be 2-D or 3- D appearance model. These models are used to match with moving regions and describe vehicle motion [15]. This method is accurate and robust but costs high. Different methods to model v e h i c l e s include multi view model and deformable template. Region Based tracking: Region based tracking detects vehicle outline as associated regions within rectangular, oval or any other simple shape which can characterized by area, coordinates, centroids ,edges or intensity histogram etc. Tracking is performed by utilizing the association between the region characteristics. The regions were matched with Kalman filtering in shape based approach [6]. Camshift algorithm is applied for video object based on histogram back projection. Feature Based Tracking: The detected features are used to perform matching in consecutive frames. The vehicle features are followed in a changed space instead of pixel space. Several techniques use combination of SIFT, HOG and Haar like features for tracking. Feature based tracking perform well in occluded conditions. The challenge is to appropriately choose features that efficiently represent the vehicles. 
Video based vehicle detection is an active area of research in intelligent transportation system. In this review, we have presented a comprehensive study of vehicle detection and approaches. We have first introduced single camera approach which includes appearance based andmotion based models for vehicle detection. We have discussed different approaches for recognition, classification and tracking. Challenges in multi-view vision system and different approaches in multi-view system have been reviewed. We hope that this paper provides substantial information regarding video based vehicle detection, recognition and tracking providing insight for researches in related area.    View More