Our objective in this chapter is to detect the motions of dynamic obstacles on the expressway to warn drivers so as to avoid possible traffic accidents. We first focus on detecting the motions of nearby moving vehicles. Some compound situations, including more than one motion occurring synchronously or asynchronously, are considers here. In the future, this system will be extended to detect other static and dynamic obstacles.

Referring to Figure 2.2, the proposed DVM is now used to detect the motions of nearby moving vehicles. The spatiotemporal information is first extracted by the method presented in Section 5.1 and is input to the STA neural module to form the attention maps. These attention maps are then divided into several overlapping blocks for extracting the local categorical features. These features are then input to the corresponding CART neural model of the blocks for motion classification. The outputs of a CART neural model in successive maps are integrated by a modified fuzzy integral technique to produce more confident classification results. Such a technique has been modified to be able to integrate the temporal information (see Section 6.2). The output of the modified technique is not a number, but a set of the highly possible motions. Finally, since all the CARTs may output their classification results, and the resulting output by a CART neural model may be more than one result, all the classification results should be collected to provide better motion detection results for drivers.

composed of two kinds of motions: lane change (right lane change and left lane change) and speed change (speed-up and slow-down). For example, to overtake a vehicle ahead, drivers may first speed up, change their lane from right to left, pass the vehicle, and then change back to the original lane. In practice, the vehicle motions are permuted sequences of these motions.

Some permuted sequences of the motions of nearby vehicles are dangerous and some are not. For example, \u201ca vehicle ahead slows down\u201d is a dangerous situation, but \u201ca vehicle ahead accelerates\u201d is not. Before detecting dangerous vehicle motions, dangerous motion analysis should be discussed first. Some motions of our vehicle itself are dangerous, including lane change and speed change. Considering the motions of only one nearby vehicle, the dangerous situations include (a) the vehicle on the right (left) lane changes lane to the left (right), (b) the vehicle ahead changes lane to the right (left) with lower speed than our vehicle, (c) the vehicle ahead slows down, and (d) the vehicle on the right (left) passes our vehicle. The closer to our vehicle another vehicle moves on the expressway, the more complex and dangerous the driving environment will be.

If there are many nearby vehicles on the freeway, the above dangerous motions could occur simultaneously because vehicles move independently of each other. For example, \u201cthe vehicle ahead slows down\u201d and \u201cthe vehicle on the right passes our

In the DVM, once an image sequence is fed into our system, the spatial and temporal information extracted by the sensory component is used to stimulate the STA neural network in the perceptual component to form the attention map. Such a process was discussed in Sections 4.1 and 4.2. Figures 6.1 and 6.2 show examples of forming attention maps, including part of the original input frames and their corresponding attention maps. These figures also illustrate the attention maps of \u201ca nearby vehicle ahead slows down\u201d and \u201ca nearby vehicle on the left changes lane to the right,\u201d respectively.

Nine dangerous motions of only one nearby vehicle are collected by the above process. These motions and their corresponding attention maps are shown in Fig. 6.3: (a) Vehicle ahead slows down. (b) Right front vehicle changes lane to the left. (c) Left front vehicle changes lane to the right. (d) Vehicle ahead changes lane to the right.